Claude Code’s first year: Make agents verify work and learn from mistakes

Boris Cherny and Cat Wu recall the early Claude Code release: a short video posted to Slack drew only a couple of reactions, and the tool handled simple engineering tasks. A year later, Cherny says, he may have many agents working or agents prompting other agents.

Cherny says that when Claude makes a mistake, he asks it to record the lesson in CLAUDE.md or create a skill rather than merely correcting that one attempt. The team also learned that verification is important, he says.

Wu asks how to make verification effective. Cherny distinguishes it from unit tests, linting and type checks, which are already straightforward to automate. For an agent, he says, the question is whether it can run and exercise the thing it built; figuring out how to do that may require deliberate setup.

Cherny recalls asking Claude, using Opus 4, to build a feature and test itself in a shell. It opened a Claude command-line session and tested the feature. He says the team now has similar loops for mobile simulators and desktop computers, but that behavior was surprising at the time.

Wu says she is working on the desktop app and uses a development skill that teaches Claude to run a local version. If the staging environment causes trouble, she asks it to check Slack for an outage or a known issue and then update the skill with what it learned.

The skill starts the desktop app and uses computer control to click through a new interface, test edge cases, fix problems and check again. Cherny says one thing he values about the team is that people across roles, including product managers, contribute code.

The speakers say the idea and the product, business, design and user context become more important when Claude can write code. Cherny recalls initially being surprised to see a designer submit a code change, then seeing that the code looked good. Wu says designers and product managers can prototype or change an app directly; finance and data-science colleagues also use Claude Code.

Asked which uses are pushing the limits, Wu points to routines. One engineer who launched voice mode set up a routine to listen for tickets, GitHub issues and bug reports; Claude prepares a fix and sends him the pull request.

Wu says the engineer also made a routine for other feedback. After she introduced a feature with an overlooked edge case, her own Claude reported that another Claude had already fixed the bug. The engineer told her his routine picks up unaddressed reports after five hours and he merges the fixes that are easy to verify.

Cherny says recurring routines made the programmatic use of Claude Code more concrete. He describes them helping with code review, following pull requests, responding to comments, fixing continuous integration and rebasing changes.

Asked about features he uses in the command line, Cherny says he used to prefer plan mode but now uses auto mode. In his experience, newer models do not need a separate planning step for his work, although some people still value the planning artifact. Auto mode lets him start one agent and move to another.

The earlier permission model asked a person to approve individual tool calls. Cherny and Wu say auto mode instead routes requests to another model for a security decision; suspicious actions can be denied and a user can later allow them. They think this helps people focus on consequential requests rather than repeatedly approving routine ones.

Cherny emphasizes continuous security work, including red teaming, penetration testing and threat modeling for attacks such as prompt injection. He says trust in that work makes it possible for him to let an agent run while attending to another.

Wu says the team first collected thousands of agent transcripts and permission prompts and had auto mode classify their safety. It then asked red teamers and internal teams to attempt prompt injections and attacks, used those attempts to create evaluations and improved the system to catch them.

Cherny says he initially doubted that routing a permission prompt through another model would work, but changed his mind after seeing the results. He describes building on models as a process of relearning assumptions from earlier engineering work.

Cherny says loops and routines represent another shift in how he works. First, he moved from writing source code to asking an agent to write it; now, he sometimes gives an instruction to a loop or routine that prompts Claude on his behalf.

Asked how engineering organizations are changing, Cherny recalls a business case about personal computers: the productivity benefit required redesigning processes around the computer, not merely putting one beside a paper-based workflow. He thinks AI adoption similarly changes processes, but expects the transition to be faster because work is already digital and Claude can use computers and code.

Cherny says Claude is embedded throughout Anthropic’s work: employees ask it onboarding questions, use it to write and review code, and use Cowork for forms. He says it frees him to spend more time generating ideas and talking with customers.

Wu asks whether the future belongs to product managers or engineers. Cherny thinks the roles are merging: he says product, developer-relations and design colleagues write code, while engineers can take an idea through implementation and work with legal, marketing and security on release.

Cherny describes moving from six terminal tabs with separate checkouts to an agent view and desktop app that manages worktrees. He says roughly half of his engineering work now happens on his phone, where he uses remote control to check agents or start a new one after a conversation sparks an idea.

On context engineering, Cherny says current models work with a minimal system prompt and tools, provided they have a way to retrieve context. Wu calls herself a context minimalist: she prefers telling the model what it needs to know and letting it find a path rather than micromanaging it with too much information.

The team is also making its supporting software leaner to leave more room for user prompts, Wu says. Cherny expects longer-running, more autonomous agents to require new interfaces, though he does not know what those will be. He expects ideas to come from the team and its community of users.