Skip to content

field notes · ai and agents

Agentic SDK integration

Agentic SDK integrationAn isometric drawing of a CI line with a Claude agent on it. A conveyor runs from left to right. On the left a pull request card waits at the start, beside a test rack of small lights. The belt passes through a CI runner arch, then under an agent station whose arm reaches over the belt; a plate on the station lists its permissions: read code yes, comment yes, production keys no, merge no. A pennant on the station shows the feature flag. In front of the belt a printer logs each step the agent takes, and at the end a team-chat sign posts when the work is done. Opening a pull request carries the card through the runner, where tests pass, then under the agent, which pins a summary and review notes and logs its steps before the card moves on and the chat sign posts. Failing a test lights one test in vermilion, and the agent pins a likely cause instead. With the feature flag off, the agent is bypassed and the card goes straight through.pull requests inon to reviewagentflag ontestsread code: yescomment: yesprod keys: nomerge: noagent logci runnerfreshcontainerpr 128team chat1234567

Pull requests ride the CI line from the left, and an agent works on them partway along. Open a pull request.

Claude agents in your CI, your code review and the tools your team uses.

Book a free first callEstimate the cost

In short

The Claude Agent SDK is Anthropic's toolkit for building agents that read code, run commands and act, on the same foundations Claude Code runs on. We use it to put agents into engineering teams' workflows: reviewing pull requests, writing release notes, sorting out failed test runs, answering questions in Slack. Zenoware is a studio in Auckland working with teams in New Zealand and overseas, and we've published open-source Claude Code tooling of our own. Start with a free first call about your pipeline.

What you get

  • A summary and review notes on each pull request

  • When tests fail, a likely cause is waiting

  • A Slack bot that knows your repo's layout

How it runs

  1. 01

    Pick the first job

    We'd look at your CI and review habits first and pick one job worth automating.

  2. 02

    Design the agent

    The agent gets only the permissions that job needs, and it logs what it does.

  3. 03

    Roll out

    It ships behind a feature flag (with a dashboard) and we walk your team through it.

  4. 04

    Tune

    Afterwards we keep tuning prompts and tools based on what people use.

Who you'll work with

Brush sketch of JP Caparas

JP Caparas

Founder and lead engineer

Thirteen years of principal engineering, including development lead at Public Trust. OpenAI Codex and Devin ambassador in New Zealand. You talk to the person who writes the code.

About the team

Questions

What is the Claude Agent SDK?

Anthropic's library for building agents on the same base as Claude Code. It runs the agent loop for you (tool calls, file access, permissions, keeping context under control) and you write the agent in TypeScript or Python. It used to be called the Claude Code SDK.

Can AI review pull requests?

Yes, though we'd call it a first reviewer, not the last one. It reads the diff and enough of the code around it to spot a likely bug, a missed test or a broken style rule, and leaves comments. A person still approves the merge. The nice part is your reviewers stop writing the same nitpick for the fortieth time.

How do you keep an AI agent in CI safe?

The boring answer is permissions. An agent in CI doesn't need your production keys, or write access to every repo, so it doesn't get them. We run it in a container that's thrown away after each job and log each tool call. And at first it only comments. If it suggests something silly, a person ignores it and nothing breaks.

What is the difference between the Claude Agent SDK and the Claude API?

The API is the model on its own. Anything agent-like, such as looping or calling tools, you build yourself. The SDK hands you what Anthropic built for Claude Code: the loop, tools for files, the shell and the web, and permissions. For most CI jobs we'd reach for the SDK, because there's less of our own code to look after.

You might also need

Tell us what isn't working.

The first call is free, and we'll tell you straight whether we're the right people for it.

Start a project