A Fireside Chat with Cat and Thariq from the Claude Code Team

Fetching the source article so the paragraphs stay factual.Simon Willison published an edited transcript and notes from a fireside chat he hosted at the **AI Engineer World's Fair** with **Cat Wu** and **Thariq Shihipar** of Anthropic's **Claude Code** team. The session covered **Claude Code**, **Claude Tag**, **Fable**, coding-agent security, evals, tool design, and how Anthropic uses the stack internally. Headline numbers from the talk: **Claude Tag** now lands **65%** of product-engineering PRs for the Claude Code team; the Claude Code system prompt was recently cut by **80%** for frontier models such as **Fable 5** and **Opus 4.8**; and Anthropic only ships features that first clear an internal bar for active use and retention among employees and early customers.

On product mechanics, **Claude Tag** is multiplayer by default in Slack: teammates can co-steer a session, the agent can run proactively (for example monitoring bug reports and opening PRs for the life of a channel), and **team memory** is a per-channel markdown file that persists preferences in natural language. **Auto mode** sits under that flow: on each turn or bash call, a **Sonnet classifier** judges the tool call against conversation context and dynamic permissions, including when a sandboxed request needs a network escape. The team has used auto mode internally since **January**, runs thousands of evals and external red teams against prompt injection and data exfiltration, and treats it as the safety layer that makes Claude Tag viable. Code review is layered the same way: humans (code owners) still approve critical surfaces such as the system prompt; a GitHub review bot covers every PR; and for outer-layer files where automated review has proven to catch issues, human review is being removed after months of incident-driven eval expansion.

For builders, the session flips several defaults. Idea-to-working-software timelines collapse from multi-month PRD cycles toward roughly a week in product domains, so **product taste and business sense** rise relative to pure execution skill (infra still needs detail). **Rewrites** are treated as viable when tests exist, because the codebase is often the only complete spec; Anthropic has already moved Claude Code onto a Bun-in-Rust stack. Prompting advice for current frontier models is leaner: fewer examples, fewer hard "don't do X" rules, more context, and wording that leaves judgment to the model (for example when to verify a front-end change). Separate system prompts per model exist for that reason; older models still carry the full prompt. Remote control—phone or browser driving a local CLI session—became a real nightly workflow after internal rollout, even though product staff did not originally prioritize it.

In market terms, Anthropic is positioning Claude Tag as the team collaborative layer above solo **Claude Code** sessions, competing with the broader agent stack (Cursor, GitHub Copilot, and similar tools referenced in the post's framing) on multiplayer work, company-context search, and proactive PR generation rather than only interactive coding. Shipping gates are retention-based dogfooding ("ant fooding"), not feature volume. Security is sold as differentiated infrastructure: Swiss-cheese defense, RL against attacks, credential injection so agents can call services like Datadog without holding secrets, trusted devices for remote control, and Claude Tag identities that can use provisioned credentials instead of acting as a human. OpenAI's own GPT-5.6 guidance similarly favors leaner prompts; Willison notes roughly **10–15%** eval gains and large token/cost cuts in OpenAI's coding-agent sample, which lines up with Anthropic's **80%** system-prompt shrink.

Practical takeaway: treat auto mode, eval-backed automated review, and public-channel dogfooding as the operating model worth copying, not only model choice. Soften absolute prompt rules, keep tool cardinality low and roles distinct, and use Claude Tag-style multiplayer sessions where social visibility reduces slop and teaches norms. Watch next for Anthropic's promised auto-mode security evals, further outer-layer human-out-of-the-loop review, and whether Claude Code system and tool prompts get published as documentation the way chat system prompts already are.

Offset for the craft side of the same talk: Thariq's advice for agent-induced skill compression is to take on more ambitious work (his example was one-shot Fable video editing with ffmpeg and Remotion, including dynamic crop and HTML slide reconstruction). That is the concrete next bar for individual engineers, not a generic industry forecast.Done — six paragraphs covering the event, product mechanics, builder impact, market context, practical next steps, and the ambition offset. Key cited facts only: **65%** Tag PRs, **80%** system-prompt cut, auto mode since January, Bun-in-Rust, Sonnet classifier, channel markdown memory.