Eliciting the tribal knowledge in people's heads that was never written down anywhere is an art and the biggest barrier to the diffusion of AI in enterprises.

This is good timely note on trustworthiness and safety of enterprise agents. As models and harnesses commoditize, we need to rethink how we can teach agents efficiently and verify their work at scale. It’s not just what an agent can access or do, but who can teach it and who can update its behavior. Our access controls and interfaces also need to evolve with this. The people teaching and updating the agent might not be the same ones verifying its work and signing off on it. The system needs to enforce that separation allow verifying the work efficiently. This is a layer that should work across models, learn from demonstrations and corrections, and provide decision trails as it evolves: what changed, why, who taught it, who verified it. It should also capture how the agent made decisions and on what basis when it did the work. Over time, through learning, it should reduce the verification burden as well. If you can verify it, you can solve it through AI. If you can verify it at scale, you can automate it through AI. Otherwise, every update to the agent becomes another thing we have to trust without being able to evaluate or verify.

Anthropic really nailed down delegation to the background threads/subagents with Opus 5.5. This ensures that the main thread doesn’t use a lot of context/tokens and stays focused on orchestration and getting the task done. This lets you get a lot out of the Max plan with Opus 5.5, as it doesn’t fill up the context window as fast. Astra so far with Codex hasn’t been able to nail this down, leading to much faster burn on the OpenAI Max plan. Curious to see how Sol 6.1 does overall with this.

"Build automatic evals", they said. "It’ll be fun", they said. There are > 300 unwritten exceptions in people's heads, and the source of truth is an Excel file called final_v7_FINAL.xlsx, they never said.

🕵️

𝗜𝘀 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗗𝗼𝗼𝗺𝗲𝗱? Anthropic’s Economic report from yesterday spells doom on software engineering. As a profession, it’s almost dead. As AI models improve, the "almost" will become imminent, and then instant. So where do all of us engineers go? Who survives the aftermath? The answer took its own sweet time to arrive for me: it’s the software engineers with "taste". When I started coding in 2012, I didn’t imagine one of the most analytical professions would be taken over by something as subjective as taste. Taste, or more crudely "good judgment", is what’s ruling the world of AI now. AI coding agents have changed the nature of software engineering work over the last year. Implementation is handled with increased accuracy. The engineer’s job is now deciding what constraints matter, and how to avoid needless complexity & cognitive overload. This judgment is pretty much the hottest commodity in tech. I faced this first-hand at my former startup, Contextual. While trying to hire a senior developer, we had to tweak our hiring process. Instead of a standard assignment that took 2 hours to solve, we gave practical questions exactly in line with the work. The interviewer was told to focus on soft skills, on how well a candidate explained. Suddenly, the journey one took to arrive at an answer became more important than the answer itself. And that's the primary skill one needs to work with an agent! This offered deeper insight into the candidate. Even though the opening was for a production engineer, we found our first sales engineer. Without the soft skills filter, the candidate would've gotten lost in product-engineer-hiring-loops. This was one example. There could be many more. But how does one judge taste in the age of AI? Do you allow AI during interviews? This question needs one foot in history & one in the future. Most interviews were designed for a world where coding was a scarce skill e.g. LeetCode’s algorithmic exercises tested unaided implementation ability, but they never represented the actual job. Engineers use AI every day. At my current startup, we realized soon that if we prohibited AI during interviews, we'd be evaluating people in an artificial environment. Interviews should mirror the real job. The question goes beyond just can you code without AI, but can you use AI effectively to solve complex problems? We’re designing exercises that can't be solved with a single prompt. They require candidates to think critically, verify AI outputs, pay heed to instructions, identify gaps & iterate. The same applies to system design: can you reason from the systems level? Since we’re building a system to capture the judgment of organizations, it becomes pivotal. Our interviews now primarily focus on how one reasons about an abstract. After all, that reasoning & critical thinking is where "taste" resides. But it needs someone with taste to decipher that. :) P. S. DM me to get a taste our hiring process!

when the same general "token capital" is available to everyone, the only real moat/differentiation is the "human capital" you compound forever inside an enterprise.

This is bigger than it seems for the AI agents. S3 Files lets you mount any S3 bucket as a native NFS on any container or lambda with ~1ms latency via EFS under the hood. Why it matters for agents: no more copying data or bridging object <-> file abstractions. Agents can now read/write S3 directly as a mounted filesystem. Multiple agents can share the same mount with close-to-open consistency. Long-term storage becomes the same as the short-term storage. Agent runtime bootstrap and teardown become trivial and instant while your data stays durable in S3 with auto bi-directional sync.

Claude Code on web/iOS is a big productivity unlock for small changes. But for bigger changes, the biggest blocker for me was not having the gh cli. Can't bring in GitHub context - no PR reviews, no issue management, no fine-grained operations. You end up pulling locally anyway which defeats the whole point. Custom environments fix this: `gh auth token` → create new environment → network access "Full" → add GH_TOKEN env var. Remote session installs gh at runtime, picks up the token. Done. GitHub-related skills now work from the browser/phone. No more local pull as an extra step.

> opened codex web > codex asks to integrate with slack > me: why not? i can delegate all things directly from slack > started a task from a slack thread > codex returns the task link > coworker opens the task link without any auth and is able to see everything > tried the link in the incognito mode > link still opens up with all of the code/details for anyone out there to see > tried for 30 min to find a setting to disable this. > found nothing > manually disable links so far one by one > disables codex in slack :( > back to claude/codex duo in tmux