Anthropic engineers described how Claude Code, Claude Tag and Fable are reshaping product development, team collaboration and security practices. This piece summarizes a fireside chat with Cat Wu and Thariq Shihipar, covering coding agent workflows, evals, tool design and how Anthropic dogfoods these systems internally.
How day-to-day work has changed
- Since Claude Code’s launch (it was first mentioned at the Sonnet 3.7 announcement and the team says it came out in February of last year), development workflows have shifted: better models and an auto mode reduced the need for constant human supervision.
- Cat Wu explained that many routine implementation tasks are now delegated to Claude, freeing engineers to focus on product experience and creative work. Fable further accelerated one-shot feature implementations for many use cases.
- Thariq Shihipar emphasised higher-quality expectations: he uses Fable to edit videos and expects outputs to meet brand standards within hours.
Which conventional software engineering practices changed?
- Prioritisation and product taste have become more valuable because the time from idea to build has shortened dramatically (from six–twelve months down to as little as a week in some cases).
- Rewrites are now acceptable: Thariq argued that rewrites can be productive if backed by good test suites; they rewrote Bun in Rust internally and have been shipping it.
- Tool design: the team prefers a low-cardinality toolset, and keeps the dedicated file-edit tool mainly because it enables a deterministic UI that shows code edits for user approval.
Claude Tag as a collaborative layer
- Claude Tag is a multiplayer Claude living in team collaboration tools (launched in Slack internally). It is proactive, supports team memory, and can continuously monitor a channel and act (for example creating PRs for bug reports and tagging the relevant engineer).
- Anthropic’s internal version of Claude Tag currently lands 65% of the product engineering PRs for the Claude Code team — a major shift in how work is split between Claude Code (complex, interactive tasks) and Claude Tag (proactive, ongoing tasks).
- Non-engineering teams use Claude Tag as a company search engine, for metrics queries, feature explanations, and coordination tasks; the tool can surface recordings or clone the codebase to show what a feature looks like.
Code review and automation
- Critical areas have code owners who must approve PRs (the system prompt is an example). A GitHub code-review bot reviews every PR and often performs the bulk of the review.
- For outer-layer changes the team increasingly relies on automated code review after months of building trust and expanding eval sets. Incident reviews feed failing PRs into evals so automated checks improve and regressions are prevented.
- The team’s long-term aim is to reduce humans-in-the-loop where safe, but for core product changes manual review remains.
Models, evals and the system prompt
- Anthropic builds eval suites so new models can be drop-in replacements: when adopting a new model they run the entire eval set to ensure the new model is at least as good as the previous one (for example comparing Fable to Opus 4.8).
- The Claude Code system prompt has been reduced by about 80% thanks to Opus 4.8 and Fable. The team found removing example-heavy, over-constraining instructions improved creativity and performance.
- The prompting approach now favors fewer hard "do not" constraints, more context, and behavioural evals to capture user-facing preferences (for example whether the model should continue working without pausing for approval).
Auto mode and safety
- Auto mode has been used internally at Anthropic since January and was first made publicly available on March 24. The team has run thousands of evals and commissioned multiple red teams to find adversarial cases.
- According to the team, they have mitigated the main risk categories they were concerned about (prompt injection, data exfiltration) and judge the residual risk to be lower than the average human reviewer for those classes of issues. They will publish evals to support these claims.
- Auto mode uses dynamic permission handling and interacts with the system sandbox. A Sonnet classifier judges tool calls and context. Claude Tag uses auto mode, which the team cites as a key enabler for safe Slack bot behaviour.
- Additional security measures include credential injection through proxies (so agents can access APIs without holding credentials directly) and trusted devices for remote control users. Claude Tag can be provisioned with its own credentials for easier auditing.
Product- and workflow-level observations
- Remote control (controlling a local Claude Code session from a phone) proved unexpectedly popular: some developers start remote sessions on their laptop and then control them from their phone while relaxing.
- Anthropic started shipping Claude Code on a Bun-in-Rust internal rollout (the team indicates they shipped it to everyone on June 17).
- Fable can competently edit video (transcription, adaptive cropping, slide substitution using HTML sources, ffmpeg/Remotion) — Thariq used it to edit a conference talk and the product launch video.
Culture, ambition and craft
- Anthropic emphasises dogfooding (internally called "ant fooding") and public internal work. Public channels and visible usage of tools help teams learn social norms for steering agents.
- Leaders encourage engineers to be more ambitious: as models automate routine tasks, humans should pursue larger, higher-impact problems.
- Product managers on the team blend engineering, design and PM skills and often step in to fill gaps quickly to keep the idea-to-customer pipeline moving.
Current limitations
- UX and design taste: models still struggle to produce delightful interaction design without human refinement.
- Real-world interaction: Thariq would like models to orchestrate scientific experiments and interact more with real-world processes.
Closing points
- Anthropic’s experience shows faster feature development and automated review are feasible but require months of infrastructure and eval investment.
- Core pillars of their approach: robust eval suites, auto mode, team memory (Claude Tag), credential-injection proxies, and heavy internal dogfooding before public release.
The full session was recorded and the team plans to publish supporting eval materials; this summary captures the main points raised by Cat Wu and Thariq Shihipar during the conversation.
