Coding agents move engineers away from manual implementation and toward product judgment, prototyping, and higher ambition.
2
Claude Tag adds a proactive, multiplayer layer to Claude Code by working in Slack, remembering channel preferences, and acting on bugs without being manually invoked each time.
3
Anthropic treats trust as an engineering process built from evals, red-teaming, code ownership, sandboxing, and gradual removal of human review.
Summary
Cat Wu and Thariq Shihipar describe how Claude Code and Claude Fable have changed their daily work at Anthropic. They spend less time watching permission prompts and writing routine implementation code, then use the recovered time to decide what is worth building and to prototype more quickly. They discuss rewrites, test suites, code review, and the internal use of Claude Tag, a proactive Slack agent that can monitor bugs, create pull requests, search company context, and retain channel memory. Safety is a large part of the discussion. Anthropic uses evals, external red teams, sandboxing, dynamic permissions, and credential injection to make Auto Mode suitable for long-running work. The speakers are candid about remaining limits, especially the design taste of current models and their ability to interact with the physical world. Their advice to engineers is to stop limiting projects to what was previously affordable and attempt more ambitious work.
Coding agents free engineers to spend more time on product decisions
Cat Wu says early Claude Code required close monitoring. She read every permission prompt carefully and often refused access. With newer models, the team delegates more menial implementation and spends more time deciding what experience to provide users. Thariq Shihipar describes a related change in his own work: because model outputs are so capable, he now expects himself to produce higher-quality work faster. He gives video editing as an example, where the result must meet Anthropic's exacting brand requirements within a few hours. The tools have raised the standard for what the engineers expect from themselves.
Product sense matters more when ideas can become software within a week
Cat Wu contrasts an older workflow with the current one. Previously, product managers might spend six months aligning with customers and cross-functional teams before writing a detailed implementation specification. Now the time from an idea to a working build can fall from six to twelve months to about a week. That makes business sense and product taste more valuable because teams can build many possible ideas quickly. Engineers need to judge what is worth building and what might affect the business. Wu adds that infrastructure work still demands careful attention to technical details.
Rewrites become practical when the codebase and tests provide a usable specification
Thariq Shihipar argues that rewrites are now a reasonable option, especially when a team has a good test suite. Rewriting forces the team to make those tests explicit. He also describes a codebase as a specification, sometimes the only complete specification available because nobody knows every branch of a large system. Simon Willison adds that agents make it possible to prototype several implementations and compare them. He now prototypes on his phone during conferences, then returns to the work later. The discussion mentions Anthropic rewriting Bun in Rust and using that version internally.
Claude Tag turns coding agents into proactive team participants
Cat Wu describes Claude Tag as an agent that lives in a team's collaboration tools, starting with Slack. It is multiplayer, so teammates can work with the same session. It is also proactive: a team can ask it to monitor bug reports in a channel, create a pull request, and tag the engineer who last changed the relevant code. Claude Tag can remember channel preferences, such as debugging outages while ignoring warnings. Anthropic's internal version currently lands 65% of product engineering pull requests. Claude Code remains better for complex interactive work, while Claude Tag handles recurring work on a team's behalf.
Anthropic uses internal adoption and retention to decide what ships
Wu says Anthropic dogfoods its products every day. When the team encounters a task the product cannot handle, it tries to fix the product rather than finding a separate workaround. Employees and early customers use features before public release and provide direct feedback. The team sets an internal bar based on active users and retention. Engineers know what they need to reach, and features that fail to retain users are not shipped. Wu gives remote control as an example of a feature whose value she initially missed. Team members now leave laptop sessions running and control Claude Code from their phones.
Trust in automated code review grows through gradual coverage and evals
Anthropic still uses code owners for important areas, including the Claude Code system prompt. Those owners manually approve changes to the areas they control. For outer layers of the codebase, Claude Code increasingly performs the review. The transition took more than six months. The team began with human review everywhere, then removed it for file areas where automated review consistently caught the issues that mattered. Incident reviews feed new cases into the review system, and those pull requests become part of an eval set so later changes do not lose coverage. The speakers present this as a process of building confidence rather than switching off human review at once.
Auto Mode combines model judgment with sandboxing and adversarial testing
Wu says almost everyone at Anthropic uses Auto Mode for long-running Claude Code work. The team has run thousands of evals and commissioned red-teamers to create environments with prompt injections and malicious inputs. Shihipar explains that a Sonnet classifier examines tool calls along with the conversation context and the user's instructions. Auto Mode can make permission decisions that depend on the current request, and it also evaluates requests to leave the network sandbox. Anthropic uses credential injection so an agent can access a service without directly seeing the underlying credential. Wu says the system does not catch everything, but the team has focused on reducing risks such as prompt injection and data exfiltration.
Frontier models need less rigid prompting and more appropriate context
The speakers say Claude Code's system prompt was reduced by 80% for newer frontier models, partly because older instructions overconstrained the model. Examples that once helped can limit a more capable model's creativity. The team now gives models more context and fewer absolute commands such as 'always verify.' Instructions need to account for edge cases, since a rule that is correct in 90% of cases can confuse the model in the remaining 10%. Different models receive different system prompts. The shorter prompt is reserved for the newest models, while older models retain more detailed instructions.
Engineers should use cheaper implementation to attempt larger projects
Shihipar acknowledges that people can feel loss when models absorb work that used to define their role. His answer is to raise ambition. He points to Jared's work writing Zig and later rewriting Bun in Rust as an example of taking on a larger challenge with Claude. Wu describes product work as filling whatever gap blocks an idea from reaching users. That can mean building a first version, improving a design, or automating launch updates. Their advice is to stop assuming that every imagined trade-off is real before testing it. Anthropic's culture encourages people to attempt the bigger version of an idea.
"It freed up a lot of our time to think about more creative work like what is the right experience that we should be providing to our users now that we know Claude Code can implement a lot of it."01:46
Who should watch
You are building coding-agent workflows and need concrete ideas for using them across a team rather than only at an individual terminal.
Your team is deciding how much human review, sandboxing, and evaluation to require before agents can make changes automatically.
You are reassessing product and engineering habits because implementation has become cheaper but choosing the right project remains difficult.
