At the AI Engineer World's Fair earlier this month, Simon Willison sat down with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team for a fireside chat covering how the team designs coding agents, thinks about memory, and dogfoods its own products. Willison has since published an annotated transcript alongside the full video.
One of the more concrete disclosures from the conversation is usage-related: Claude Tag, Anthropic's newer Slack-based collaborative agent, now handles roughly two-thirds of the product engineering pull requests generated for the Claude Code team itself. Wu described Claude Tag as built to be multiplayer by default and proactive rather than reactive, meaning it can watch a channel for signals like bug reports and open a pull request, tagging the relevant engineer, without being explicitly asked each time.
Memory design came up as a genuinely unsettled area. Shihipar explained that Claude Tag currently keeps memory scoped per Slack channel, with every instance of Claude operating in that channel sharing a single memory store backed by a markdown file. Individual sessions can run independently but are able to write back into that shared channel memory, a structure the team arrived at after what Shihipar described as ongoing internal experimentation, since the right way to handle agent memory is often unintuitive.
On prompting, the pair offered a piece of practical advice that runs against a common instinct among developers: stop overloading prompts with long lists of examples and explicit instructions about what not to do. Both said Anthropic's newer Fable model performs better with leaner prompts, and noted that Claude Code's own internal system prompt has shrunk by roughly 80 percent over recent iterations as the underlying models have grown more capable of inferring intent without heavy scaffolding.
The conversation also touched on how the team's relationship with the tool has changed since Claude Code launched a little over a year ago. In the earliest period, Shihipar said, engineers had to closely monitor every action the agent took, reviewing permission prompts one by one. With each subsequent model generation, the team has been able to step back and delegate more of the routine implementation work, freeing up time for higher-level decisions such as user experience design.
The pair also discussed security considerations specific to coding agents, evaluation methodology, and tool design choices that shape how reliably an agent can be trusted with broader autonomy.
For developers watching how frontier labs build their own internal tooling, the exchange offers a rare look at Anthropic treating its own products the way any engineering team would: as something to be measured, second-guessed, and iterated on, rather than treated as a finished artifact.