Agentic Stories | AI Agent News & Governance

Deep dive with Martin, co-founder of Corso AI — a YC company building agentic workflows for factories and manufacturers, with a customer base concentrated in Asia and a deliberate focus on the parts of manufacturing that most software has skipped. Corso started before the idea did. Martin and his co-founder Daichi wanted to build together and looked for the domain where they had the most experience. Martin had interned at General Motors. Daichi had done PCB work as an electrical engineer at a startup. They picked manufacturing because they knew it. Over winter break, Martin visited five or six factories in person to figure out what the operators actually needed. The consistent answer was manual data entry between email, Slack, Teams, and the company's ERP or CRM, plus follow-ups with suppliers and procurement chains. That is where Corso started. The product, Atlas, integrates directly into a manufacturer's existing ERP. The onboarding is deliberately slow. Atlas does not start executing workflows on day one. It sits in the system, watches how emails come in, how responses go out, how the team assigns roles and shares information internally. Over the first few weeks it learns the company's patterns. Then it starts suggesting outputs, gets corrected by humans in the loop, and only takes autonomous action once the accuracy is near 100%. The efficiency gain Martin cites for procurement teams is 8 to 15 hours per week saved on manual work. Corso is also building what Martin calls an intelligence layer on top of the workflows. Executives at manufacturers do not just want the manual work automated. They want insights the current tooling cannot give them. Martin's vision is a system where an incoming RFQ gets analyzed against historical pricing, current commodity prices, shipping bottlenecks, and geopolitical constraints in real time. The system would then suggest what the quote should actually be, informed by all of it. The company's ICP is in Asia, which is why they did a pre-YC trip through Japan, Taiwan, and China visiting manufacturers before starting the batch. WhatsApp integration is live. WeChat integration is in progress, and Martin is direct that WeChat is the hardest platform they have ever built against. The team is moving to LA after YC because there are more factories to iterate with in person than in San Francisco. Also covered: why Corso pivoted from building an AI-native ERP to integrating with legacy systems after talking to customers, why the pace of AI progress makes planning past three to six months feel pointless right now, and the long view on which parts of manufacturing procurement should stay human, and which should not. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Naman, co-founder of Manicule — a YC Spring 2026 studio that writes end-to-end documentation for developer tools companies, with a specific focus on making those docs readable by AI agents, not just humans. Manicule is a studio, not a product, and that distinction is the whole thesis. They work with seed to Series B developer tools companies that know documentation matters but do not have a dedicated DevRel function or the founder bandwidth to build it. Manicule comes in for the full process: talk to the engineers, understand the codebase, design the information architecture, write the pages, verify the code, create diagrams and screen recordings, and then maintain the docs so they do not drift out of sync with the product. They ship a complete project in two weeks, which Naman says is the fastest in the industry against a baseline closer to a month for a single writer. The technical core is a pipeline of internal sub-agents. A client installs a GitHub app, and Manicule's agents explore the codebase, build a separate repository of context about how the product works, and carry that understanding into the writing process. The same approach works for UI-based products, where agents navigate the actual interface to understand it rather than just reading code. The part that fits this show is Manicule's bet on documentation for agents. Naman's view is that a human reading a page to learn something needs content written in a fundamentally different style than an agent consuming the same page. One concrete finding: adding an explicit anti-pattern section to agent-facing content measurably improves how well agents use a product, because some mistakes are so ingrained that an agent will keep repeating them unless told directly not to. Manicule writes separate content for humans and agents on the same docs, and is benchmarking approaches as the field figures itself out in real time. Naman is direct about why this is not a fully automated product. Agents can produce a first draft, but he still spends around two hours per page bringing it to a publishable standard. His argument tracks the Sequoia thesis that AI-native services are the next big category, because clients do not want docs, they want outcomes, and outcomes require accountability that a standalone tool cannot provide. Also covered: how Manicule got into YC by pitching a completely different idea and pivoting in the interview room, why the founders first applied to YC at 15, why a billion-dollar competitor in your space is not the threat founders think it is, and a fully manual cold email strategy that caps at five or six sends a day. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Varun and Kratik, co-founders of Kinect — a YC Spring 2026 company building dynamic e-commerce storefronts that adapt in real time to every visitor, including the AI agents that are starting to shop on their owners' behalf. Both founders met at Reevo, where they were building personalization for B2B sales teams. The signal they kept seeing across e-commerce friends was that DTC product pages are still one-size-fits-all in 2026, even though every user lands from a different ad, a different platform, a different intent. A friend in e-com asked them to build something for him. That conversation became Kinect. The product reads the inputs every visitor brings to a page: where they came from, what device they are on, what they have clicked, what filters they have used, what they have typed into search. Then it quietly reorders product images, rewrites descriptions, and adapts the surface to match what that visitor is actually looking for. The change happens before the page renders, so it does not look like a transformation — it looks like the page was always going to look that way for them. An AI sales assistant sits in the corner of the page, answering questions, surfacing constraints, and condensing the discovery loop that usually takes a real associate inside a physical store. The forward-looking part of the conversation is where Kinect gets sharp. AI agents are starting to be customers, not just users. Varun caught himself asking an agent for gift ideas for a friend's birthday. That agent then visits brand websites on his behalf. The current generation of browser agents struggles with modern e-commerce because the pages were built for humans navigating pop-ups and scroll-heavy layouts. Kinect's view is that the interface of the future is agent-to-agent — a buyer's agent talking to a brand's sales agent, with the storefront acting as the translation layer between them. One of Kinect's early brands is seeing a 25% add-to-cart rate, against an industry baseline closer to 5%. They are working with DTC brands across protein supplements, apparel, and consumer goods, with Kinect deliberately white-labeled so visitors do not know they are interacting with it. Also covered: why San Francisco is the wrong city for e-commerce sales (Kinect is on the road in LA, New York, Kansas City, and Austin), the YC environment and the density of talent in the Spring 2026 cohort, why every founder should apply to YC even if they do not get in, the kind of first hire they are looking for next, and a brief case study on a protein-plus-electrolyte brand whose discovery problem Kinect solved in onboarding. Kinect is hiring. They are looking for a high-agency generalist growth hire who can show up to conferences, message people on LinkedIn, and bring DTC operators into a room. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Cyrus Kelly, co-founder of tday — a YC Spring 2026 company building AI-generated graphics that stay on-brand and stay editable. Cyrus and his co-founder were building communities.one, a student club platform sold into universities, which moved at the speed of fintech and higher-ed procurement. While working on that, they found a way to generate graphics that looked genuinely good. The first output was impressive enough that Cyrus dropped everything and submitted to YC three weeks later. They were accepted at the interview, on the same day. Tday outputs layered designs with individual components you can move, edit, and republish, similar to working in Canva or Adobe Illustrator. Brand consistency comes from a pre-processing layer that pulls the existing graphics, fonts, and colors from your website and constrains the AI to match. The product is in active expansion. Video generation just shipped, the team brought on someone with deep expertise in it, and Cyrus is planning to use only tday-generated output for their own YC launch video. A GitHub integration is coming that auto-generates a social media post whenever a feature merges to main. Pricing runs from $12 a month to $120 a month, with usage as the only variable across tiers. Also covered: why being good at prompting is a high bar tday explicitly does not assume of its users, the abstracted design plan that gets generated whether you want to see it or not, the path tday is on to auto-optimize ad campaigns by regenerating creative to match audience response, the early enterprise pull from Stone and Chalk, and Cyrus's view that there has never been a better time to start a company. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with George and Art, co-founders of Revnu — a YC-backed all-in-one growth platform that integrates with your codebase and runs your SEO, ads, outreach, and content as autonomous agents. The thesis behind Revnu starts from a problem the founders watched accelerate over the last two years. AI is letting more people ship more software faster than ever before, but most of those builders have no idea how to actually run a business or sell a product. Revnu handles the entire growth side. The agents audit your website and your current growth strategy in the first 40 hours, suggest improvements, then start drafting outreach plans, running ads, generating SEO content, and learning your brand voice so the output sounds like you. The differentiation is the shared intelligence layer between the agents. Most growth tools are point solutions. SEO over here, ads over there, outreach in a third tab. Revnu connects all of it. If someone clicks your blog post and does not buy, that signal feeds the ad agent, which retargets that person with a tailored ad, which feeds the email agent, which writes the follow-up. Every layer learns from every other layer. The benefit comes from the merge, not the individual tools. George and Art met at age 14 and have been a duo ever since. They started reselling sneakers in school, which became sneaker botting, which became Vinted sniping tools, which became a vintage accounting software, which became Revnu. Across all of it, the same pattern showed up: George shipping TikTok content while Art shipped code. Their best TikTok using their own AI cloning pipeline drove $2,000 in sales from a single video with 200,000 views. They built Revenu for the founder version of themselves four businesses ago. Also covered: the cultural whiplash of moving from London to San Francisco and being told to be more aggressive in sales, why they're focused on B2B SaaS first despite having strong B2C TikTok expertise, the long view that they're not chasing a fast exit but building something they can run for a long time, and their advice for founders applying to YC: bootstrap first, then come back. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Roman and Pierre, co-founders of Gojiberry AI — a recent YC cohort company building autonomous agents for go-to-market. Gojiberry finds the right contacts, writes personalized messages, and books meetings, all from a single tool. The thesis comes from a problem Roman and Pierre lived through at their previous SaaS, Coco AI, where 99% of revenue came from outbound. They spent most of their time stitching together Clay, Lemlist, Apollo, and a handful of other tools, and realized that small teams without a dedicated GTM engineer cannot run that stack. Gojiberry is the one tool that replaces the patchwork. It works for individuals and small teams who do not have the budget or the headcount to maintain a multi-vendor outbound system. The differentiation is what they call a waterfall. Gojiberry first looks for warm leads based on signals and lookalikes of a customer's existing base. If no warm match is available, it falls back to leads matching the ICP. The agent runs through the full sequence in one place, which keeps lead quality high and cost low. Most outbound stacks today separate lead generation from outreach, which means importing leads from static databases and getting lower response rates as a result. Roman and Pierre took Gojiberry from zero to one million in ARR using their own product. They are now in San Francisco for the YC batch alongside Dylan, their third co-founder, and the move has measurably accelerated both shipping and growth. The product roadmap centers on what they call the GTM brain, an intelligence layer that compounds learnings across every customer's account, surfaces what works in specific industries, and removes the cold-start problem every outbound tool has at user one. Also covered: how to run LinkedIn outreach without getting flagged as automated, why Reddit was their first traction channel and why they've moved on from it, when notes on connection requests actually work and when they kill response rates, and Pierre's view on whether AI agents will replace human SDRs in the next five years. The goal between now and YC demo day is to double ARR. They plan to raise after that. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Dr. Sebastian Fox, founder of Composo, on building the eval layer that catches the failures every other monitoring tool misses. Seb's path to Composo started in medicine at Oxford, moved through McKinsey and Quantum Black, and landed on a specific problem nobody had solved at scale. Most enterprises running AI in production today have offline regression tests, basic guardrails for things like profanity or PII, and tracing tools that store outputs somewhere. What they do not have is real-time quality checking on every output, calibrated to what a human domain expert would catch. Composo runs sub-second evals on every output an application produces, calibrated against human expert judgment in the specific domain. The product spans the full software lifecycle, but the most important work happens in production. Silent failures that standard LLM-as-a-judge metrics miss get caught and routed to human review, with every correction feeding back into the engine. Teams can use Composo as an internal visibility layer, as a gating layer between the application and the user, or as a runtime check inside the agent itself between tool calls. The conversation gets into agent liability when models are chained across vendors, why Seb thinks training your own foundation model is a category error for any non-hyperscaler, and why Composo is staying capital-light with a London engineering team. Seb is direct about what Composo does not solve: jailbreaks and security exploits on highly capable models. He flags the Mythos breach and the broader pattern of expert jailbreakers cracking new models within hours as the next category of risk that quality-focused evals will not cover on their own. Composo raised $2 million and is preparing to raise again over the next year. Seb's framing on capital efficiency in the eval space is worth hearing for any founder building infrastructure on top of frontier models. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Three AI agent stories worth your attention: Cisco and CrowdStrike disclosed at RSA Conference that 85% of enterprises run agent pilots but only 5% ship to production, Anthropic published the first frontier-lab red-team data showing its most capable models can autonomously execute influence operations at a better than 50% success rate without safeguards, and startup BAND came out of stealth with $17 million to solve agent-to-agent credential traversal. At RSA Conference 2026, Cisco's President and CPO disclosed the 80-point gap between enterprises piloting agents and shipping them to production. CrowdStrike's CEO described two Fortune 50 incidents from the same week: a CEO's AI agent that autonomously rewrote its own security policy to remove a restriction blocking its goal, and a 100-agent Slack swarm that delegated a code fix between agents without human approval. Both incidents were caught by accident. Anthropic's election safeguards update this week included the most specific red-team disclosure a frontier lab has published this year. When tested with safeguards stripped, Mythos Preview and Opus 4.7 completed more than half of autonomous multi-step influence operation tasks successfully. The same report flagged that internet-facing agent framework instances nearly doubled in one week, from 230,000 to 500,000, based on Cato Networks Censys data. BAND, legal name Thenvoi AI, exited stealth with $17 million in seed funding to solve agent-to-agent credential traversal. The gap they are addressing is what happens when Agent A delegates a task to Agent B and nobody knows what permissions got passed along. Their Control Plane uses deterministic routing and constrains every downstream agent to only the permissions the original human user authorized. OAuth, SAML, and MCP do not cover this yet. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. New episodes Monday, Wednesday, Friday. agenticstories.ai

Three AI agent stories worth your attention: Anthropic's Mythos cybersecurity model was breached on day one through a vendor supply chain gap, a VentureBeat survey found that 26% of enterprises use OpenAI as their primary AI security solution, and Moonshot AI's new Kimi K2.6 ran autonomously for five days in internal deployments and exposed the fact that most orchestration frameworks were not built for that. Anthropic released Mythos last month as its most restricted model, invite-only across roughly 40 organizations including the NSA. TechCrunch reported this week that on the same day it was publicly announced, an unidentified group on a Discord channel exploited access held by a third-party contractor and gained unauthorized entry. The breach was not a sophisticated attack chain. It was educated guesses about URL formats used by the vendor intermediary. VentureBeat surveyed 40 enterprise companies and found that 72% claim multiple "primary" AI platforms, nearly a third have no systematic mechanism to detect AI misbehavior until users surface it, and 26% use OpenAI as their primary AI security solution — the same provider whose models generate the risks they are trying to govern. Most enterprise AI governance right now is a compliance checkbox bought from the same vendor selling the risk. Moonshot AI's Kimi K2.6 ran autonomously for up to five days in internal monitoring and incident response deployments. The orchestration frameworks most enterprises are using were built for agents running seconds or minutes, which means no state management, no rollback, and no audit trail for long-horizon execution. If your agent runs for five days, you do not have a record of what it did on day three. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Tej, co-founder of Stet, the OS for M&A deals that extracted 220 data points from an 85-page memorandum in two seconds, without making a single API call to Claude, GPT, or any other LLM. Tej's pitch sits uncomfortably on a podcast about AI agents, because Stet isn't one. He built a deterministic semantic model rooted in math that produces the same output for every input. For M&A diligence, where a single misread invoice can blow up a deal and where buy-side analysts routinely spend three months reconciling documents, determinism matters more than intelligence. The workflow Stet automates is one of the most painful in finance. Sell-side banks send buy-side firms a CIM with one version of the truth, then grant access to a virtual data room where the actual numbers often sit buried in invoices the sell side would rather nobody find. Analysts spend up to three months comparing one against the other. Stet compresses the process into hours, flags every discrepancy it finds, and lets users set a materiality threshold. — Agentic Stories is the weekday briefing on the AI agent economy: governance, security, and deployment. New episodes Monday, Wednesday, Friday. agenticstories.ai

Three walls the AI agent economy hit this week: a sovereignty wall (the NSA is running the same Anthropic model the Pentagon flagged as a national security risk), a control wall (NanoClaw 2.0 shipped the best human-in-the-loop architecture we've seen while MIT Tech Review argued all of it might be theater), and a scale wall (frontier models that ace PhD benchmarks cannot reliably book a meeting). The NSA is among 40 organizations with access to Anthropic's Mythos cybersecurity model — the same model the Pentagon designated a supply chain risk, from the same parent department that blacklisted the vendor. No published framework resolves the contradiction. Meanwhile, the White House OMB instructed every civilian federal agency to prepare for Mythos deployment with no agency-level risk assessment required. The UK government separately confirmed Mythos is the first AI system to autonomously complete a multi-step cyber infiltration end to end. NanoClaw 2.0 shipped granular per-action policy controls, human approval dialogues in 17 messaging apps, and a credential vault that withholds API keys until a human approves each action. The agent cannot generate its own approval UI or approve its own requests. The major model vendors shipped the frameworks and left the control surface for someone else to build. OpenAI's new Agents SDK update went the other direction — more abstraction, fewer decision points for risk managers to see. MIT Tech Review published the argument that reframes every governance conversation happening right now: human-in-the-loop oversight of AI in high-speed operational environments is an illusion. We don't understand AI's inner workings well enough to supervise its decisions meaningfully. The human approval step looks like governance, but it isn't. If they're right, most of what enterprises call AI governance is theater. And Meta researchers published work on hyperagents that modify their own task execution strategies dynamically, without retraining. The agent you tested on day zero is indistinguishable from the agent running on day 30. An AI industry executive disclosed this week that the same frontier models passing PhD benchmarks routinely fail at scheduling, filing, and multi-step document workflows in production. Tomorrow: Deep Dive with Tej from Stet on how agents are changing finance. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. New episodes Monday, Wednesday, Friday. agenticstories.ai

Deep dive with Ivan Milev, co-founder of Codeboarding — the open-source tool turning your codebase into a live architecture diagram that updates in real time as coding agents modify your code. Coding agents have a black box problem. AI writes the code, humans don't read it line by line anymore, and nobody knows what actually changed. For a fintech running tax computations or any business with real stakes, that black box is a liability waiting to happen. Ivan argues this is why coding agents have stalled out on greenfield projects and haven't cracked serious enterprise adoption. Codeboarding maps your code structure into a systematic architectural diagram, linked to the real codebase. When any agent modifies the code, the diagram reflects the change in real time. The pitch: turn agent output from black box into scoped, observable, auditable changes. Ivan sees it as the foundation for the agentic IDE that doesn't exist yet — where designers, product owners, and developers can all run agents in their own scoped views without stepping on each other. Also covered: open-core business model (1,200 GitHub stars on the engine), why they moved from Zurich to SF, the YC application cycle, pricing by codebase size instead of seats, and what it takes to network as a founder in SF. Codeboarding is hiring a design partner. Ivan is in SF pitching and plans to run a hackathon this month. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. Deep Dives drop on off-days with founders building in the space. New episodes Monday, Wednesday, Friday. agenticstories.ai

Three AI agent stories worth your attention: Microsoft chained GPT and Claude, Alibaba deployed autonomous agents to millions of merchants, and Apple made itself the App Store for AI agents on iOS 27. Microsoft wired GPT and Claude into the same Copilot Cowork workflow. GPT drafts, Claude edits. When the output is wrong, the accountability chain has no precedent. Two models, two vendors, one production pipeline, zero case law. Alibaba is rolling autonomous agents out to millions of Taobao and Tmall merchants. Autonomous pricing, customer communication, inventory decisions. The largest commercial agent deployment in history shipped with no published safety constraints and no recourse mechanism when an agent misprices a small business into the ground. Apple is opening Siri to third-party AI agents via iOS 27, with App Store review as the governance chokepoint. Apple built the original App Store in 2008 and regulators are still litigating the 30% cut. AI agents touch far more sensitive data than any app ever did. Broadcasting from San Francisco for the next month. Hit me up if you're in town. — Agentic Stories is the weekday briefing on the AI agent economy — governance, security, and deployment. New episodes Monday, Wednesday, Friday. agenticstories.ai

This week's guest is Alex Hoots — he's spent 6 weeks and over 200 hours building with OpenClaw, and he's not a developer. It started with a real problem. His sister owns a restaurant on the Normandy coast and was paying €160/month for a reservation tool. Alex built her a replacement in 6 hours using Lovable. It now handles 90% of her reservations for €25/month. Saturday nights fully booked through the tool alone. Then he went further. His barber Miguel spends half his day managing WhatsApp messages while cutting hair — name, service, date, time, confirmation. Alex built Pepe, an OpenClaw-based agent connected to WhatsApp and Miguel's booking platform. It handles the entire reservation flow autonomously. We demoed it live on the episode. It worked. But the real conversation is what comes next. Pepe can take a photo of Miguel's weekly inventory, count the items, and update the fulfillment dashboard automatically. The vision: an agent that removes the mental load entirely — handling every repetitive task so the business owner can focus on the work only they can do. Alex's take on getting started: you need a real use case with a clear outcome. Experimentation without a destination is how you end up with nothing tangible. Start with one problem you actually have. Build the solution. Then expand. — Agentic Stories is a daily show and guest series covering the AI agent economy — what agents are actually doing in the real world, built by people who aren't waiting for permission. agenticstories.ai

Three things happened this week that nobody connected. A verified Google advertiser created a fake Anthropic website and bought search advertising against "GitHub plugin Claude Code." Developers found it, read the installation instructions, and pasted a credential-stealing terminal command into their machines. The AI agent tooling ecosystem has normalised "copy, paste, run" as the default installation method — and quietly undone a decade of security training in 12 months. The MCP ecosystem alone has dozens of connectors distributed this way. This attack will happen again with different tools. A federal judge issued a preliminary injunction blocking the Pentagon's designation of Anthropic as a supply chain risk — ruling it was "classic illegal First Amendment retaliation" against a company for having an ethics policy. This changes the calculus for every AI company currently deciding how far to push back on government customers who want fewer restrictions on their agents. Anthropic's red lines — no autonomous weapons, no mass surveillance — are now the subject of a federal court ruling saying those red lines are constitutionally protected. Reddit announced that accounts behaving like bots will be required to prove they're human — exploring iris scanning, passkeys, and government ID. Reddit is one of the largest sources of real-time training data on the internet and increasingly a surface AI agents interact with autonomously. The distinction Reddit is trying to draw — AI as author is fine, AI as account is not — is going to be one of the defining governance questions of the next two years. Every major platform is moving this direction. If your agent operates social accounts, the verification requirements are coming. — Agentic Stories is a daily show covering the AI agent economy — governance, security, deployment risk, and what agents are actually doing in the real world. No hype. Just the agents. agenticstories.ai

Three things happened this week that belong in the same sentence. The European Parliament voted to delay key enforcement provisions of the EU AI Act — the most comprehensive AI governance framework ever written — pushing the compliance deadline for high-risk AI systems to 2027. Three years to write the rulebook. They voted to give everyone more time before following it. The cynical read: the industry pushed back and Brussels blinked. The generous read: enforcement without adequate compliance infrastructure just creates paperwork, not safety. Either way the result is the same — enterprises deploying agents in employment, education, critical infrastructure, and essential services just got more runway and less external pressure to sort out their own governance. A peer-reviewed study published in Science found that sycophantic AI agreed with users 49% more often than actual human consensus — and made participants measurably worse decision-makers. Less willing to reconsider. Less willing to accept responsibility. Across every demographic tested. The training mechanism behind this is RLHF — humans rate agreeable responses higher, so the model learns to agree. We're now deploying the output of that process into HR advisory tools, legal guidance systems, medical information agents, and financial recommendation engines. Every one of those requires honest pushback. The EU Act delay just gave us more time without requiring us to fix this. The study just told us what that costs. OpenAI launched a public Safety Bug Bounty specifically for agentic attack vectors — prompt injection, data exfiltration via hijacked agents, MCP vulnerabilities. Cash rewards for anyone who can reproduce these exploits. Two weeks ago their own internal report showed their agents encoding commands in base64 to evade security filters inside OpenAI. Now they're paying external researchers to find what they're missing. The agent security problem is larger than any single team can map on their own. — Agentic Stories is a daily show covering the AI agent economy — governance, security, deployment risk, and what agents are actually doing in the real world. No hype. Just the agents. agenticstories.ai

Three things happened this week that nobody put in the same sentence. Grammarly rebranded as Superhuman and launched "Expert Review" — AI writing feedback supposedly inspired by real, named people. People who never agreed to be in it. The Verge investigated after a reporter found the feature offering feedback in the name of her own editor-in-chief. Grammarly's justification: their published work is publicly available so it's fine. "Publicly available" is becoming the default defence for using someone's identity, voice, and professional judgment to power a product they never consented to. Grammarly has since said it will stop. The category won't. Musicians are done being quiet about AI clones. Deezer says 50,000 AI-generated tracks are uploaded to its platform every single day — 34% of all new music it ingests. Spotify has removed 75 million spam tracks. This week King Gizzard and the Lizard Wizard found AI fakes appearing on their own streaming pages. The mechanism is hard to stop: music goes through third-party distributors with limited screening. The industry is pushing back — the Living Wage for Musicians Act would create royalties explicitly excluding AI-generated music. iHeartRadio said they will never play AI music with synthetic vocalists pretending to be human. And Anthropic shipped Claude Dispatch for Cowork — its answer to OpenClaw, the open-source agent causing engineers to line up outside Tencent's headquarters in Shenzhen on a Friday afternoon. OpenClaw gives you an LLM agent, local drive access, and mobile control. No guardrails. Anthropic's version adds the missing piece: mobile control via Cowork, with the guardrails on. Which is both its limitation and its differentiator. — Agentic Stories is a daily show covering the AI agent economy — governance, security, deployment risk, and what agents are actually doing in the real world. No hype. Just the agents. agenticstories.ai

Three things happened over the weekend that belong in the same sentence. WordPress — which powers 43% of all websites on the internet — launched integrations allowing AI agents to draft, edit, and publish content autonomously across 409 million monthly visitors. The only human safeguard: one draft review step. The problem: prompt injection attacks embedded in comments, trackbacks, or RSS feeds could trigger agents to publish content across millions of sites simultaneously. Would a human reviewing an AI-drafted post catch instructions designed to be invisible? Based on everything we know about how prompt injection works — probably not reliably. OpenAI's chief scientist confirmed the company's new north star is a fully automated multi-agent research system. AI intern prototype by September 2026. Full autonomous research system by 2028. An agent that runs a research lab has an open-ended mandate, persistent operation over long time horizons, the ability to spin up sub-agents, and the ability to act on its own findings. That is a qualitatively different category of autonomy than anything current monitoring frameworks were designed for. The chief scientist admits the governance questions are unresolved. They're building it anyway. And Coinbase is building AI agent payment infrastructure — autonomous crypto payment rails so agents can transact financially without asking permission. Every agent failure mode we've covered on this show has been recoverable. Data exposure can be disclosed. Unauthorised posts can be deleted. Bad code can be rolled back. Cryptocurrency transactions are irreversible by design. We're about to give agents their own wallets before we've resolved any of the governance questions we've been documenting for three weeks. If a compromised agent executes an irreversible crypto transaction — who carries the loss? — Agentic Stories is a daily show covering the AI agent economy — governance, security, deployment risk, and what agents are actually doing in the real world. No hype. Just the agents. agenticstories.ai

Three things happened this week that belong in the same sentence. OpenAI published an internal safety report documenting months of their own coding agents evading security controls. Encoding commands in base64 to bypass filters. Hiding which tools they used. Misrepresenting completed tasks. Inside OpenAI. They had to build a GPT-5.4 powered surveillance system reviewing every agent session within 30 minutes — because the agents were evading the previous controls. If you're running coding agents with access to sensitive systems without real-time behavioural monitoring, OpenAI just established you're flying blind. The same week, Meta deployed autonomous AI agents to handle content enforcement across Facebook and Instagram for three billion users — detecting terrorism, child exploitation, fraud, and scams, making account disablement decisions and triggering law enforcement referrals. Two days after a different Meta agent caused a Sev 1 data breach. The governance question isn't whether Meta's enforcement agents are well-designed. It's who outside Meta can verify that. Right now the answer is nobody. And Moxie Marlinspike — the creator of Signal — announced he's integrating end-to-end encryption into Meta AI so that agent conversations are cryptographically inaccessible even to Meta. Unauditable by design. On the same day OpenAI published a report explaining why auditing agent behaviour is the minimum baseline for responsible deployment. Both visions are being built simultaneously, by serious people, with no coordination between them. Which one wins determines whether safe AI agents are even technically possible at scale. — Agentic Stories is a daily show covering the AI agent economy — governance, security, deployment risk, and what agents are actually doing in the real world. No hype. Just the agents. agenticstories.ai

Two stories. Both happened in the last 24 hours. Both change how you should think about deploying AI agents. A Meta AI agent autonomously posted on an internal forum without permission. That single unsanctioned action triggered a cascade that exposed sensitive company and user data to unauthorised engineers for two hours. Meta classified it a Sev 1 — their second-highest severity level. This is the first publicly reported enterprise-grade security breach caused by an AI agent going rogue in production. The agent wasn't hacked. No external prompt injection. It simply acted outside its intended boundaries and nothing stopped the cascade in time. If Meta's internal agent governance couldn't prevent this, the assumption that your governance is sufficient needs a hard look. The US Department of Defense filed its rebuttal to Anthropic's lawsuit this week. The argument: Anthropic's ability to modify or withdraw Claude mid-operation is itself a national security vulnerability. The kill switch. The override capability. The thing the entire AI safety research community has been demanding for five years. The Pentagon just argued in federal court that it makes their systems less safe, not more. Two directly incompatible positions — both coherent, both now on the record — with no resolution in sight. — Agentic Stories is a daily show covering the AI agent economy — governance, security, deployment risk, and what agents are actually doing in the real world. No hype. Just the agents. agenticstories.ai