2. Adapt It to Your Team
The library wrote this agent for any team. In this page, you make three changes to Release Readiness. You limit its tools, you set its model, and you add your team's rules to its instructions. You make the changes on a branch. Other people and triggers keep the version on main until you are sure.
Step 1: Make a branch
Click the branch button at the top of the Agent Studio sidebar, and select New branch from main. In Branch name, type adapt-release-readiness, and click Create branch. From the CLI:
autoheal branches create adapt-release-readiness
Step 2: Limit the tools
Release Readiness only reads, but its filters still keep ask_user_input and every read tool of every Integration Connection. A run that a trigger starts has nobody to answer a question, and this agent does not need Slack. In agents/release-readiness/agent.yaml, replace the two filters with these lines:
system_tools: operation_type == "read" && name != "ask_user_input"
integration_tools: "false"
name != "ask_user_input"removes the tool that pauses a run for a question."false", with quotes, gives the agent no integration tools. Without quotes, the push fails.- A subagent can use only the tools that both its own filters and the filters of its caller keep. You do not have to change the subagents.
If your project has GitHub or Datadog, keep the reads that the agent needs, for example integration_tools: integration_type in ["github", "datadog"] && operation_type == "read".
Step 3: Set the model
The agent and its subagents each name a model. In agent.yaml, change thinking: high to thinking: medium:
model:
name: claude-sonnet-5
max_tokens: 8192
thinking: medium
Change one thing at a time when you tune a model. To use a different model, write a name that your organization offers. autoheal agents validate lists the names if you write one that does not exist. Not every model accepts thinking. See chapter 2.
Step 4: Add your team's rules
Add this section at the end of agents/release-readiness/instructions.md:
## Team rules
- For each pull request, clone the repository with the `git_clone` tool and read the changed files in full, not only the diff.
- Check that each time value in the change uses the unit that the called function expects.
- Our rollback plan for every service is: revert the pull request and deploy again.
- Only read. Do not comment on pull requests and do not post messages.
Add your own rules below the library's instructions, in their own section. When you copy a newer version from the library, you then see your changes in one place.
Step 5: Push to the branch and check the tools
autoheal agents validate agents/release-readiness
autoheal agents push agents/release-readiness --branch adapt-release-readiness -m "Adapt release-readiness to our team"
The push prints the tools that the new version adds or removes. Check that ask_user_input and the Slack tools are in the removed list. Then look at the full list:
autoheal agents tools release-readiness --branch adapt-release-readiness
Check that ask_user_input shows Removed by system_tools in agent.yaml, and that the line INTEGRATION TOOLS integration_tools: false has no available tools.
In the web app, select the branch, make the same changes in the editor, and commit. Agent's tools in the status bar shows the list for the version committed on the branch.
Step 6: Compare the two versions
Run the agent on the branch with the same prompt as on page 1:
autoheal run release-readiness "This release has one change: https://github.com/autohealai/astronomy-shop-template/pull/1. Decide ship or hold." --commit <commit> --wait
Use the commit that the push printed. In the web app, run the agent while the branch button shows adapt-release-readiness.
Compare the run with the run from page 1:
- Check that the new run names the unit problem: the code gives a time in milliseconds to a call that expects seconds.
- Check that the rollback gate now uses your rollback plan.
- In the Transcript tab, check that the agent cloned the repository and read the changed files.
Step 7: Keep the change
Branches cannot be merged yet. When the branch version is correct, make the same changes on main and commit them. From the CLI, push the same folder without --branch:
autoheal agents push agents/release-readiness -m "Adapt release-readiness to our team"
Then add a trigger if you want the agent to run by itself. See chapter 5.
What to know
- Make safe changes first: narrower filters, a different model, and new rules at the end of the instructions. Change the subagents, skills and
returnsonly when you know why the library wrote them as it did. - A filter limits tools. It does not limit the shell. Say in the instructions what the agent must never change.
- The agent's History tab shows each commit to the agent, with its author. A run shows the commit that it used.
- To remove an eval or a skill that you do not need, remove its line from
evalsorskills, and remove its file. Then push again.
From the CLI
autoheal branches create adapt-release-readiness
# change agent.yaml and instructions.md
autoheal agents validate agents/release-readiness
autoheal agents push agents/release-readiness --branch adapt-release-readiness -m "Adapt release-readiness to our team"
autoheal agents tools release-readiness --branch adapt-release-readiness
autoheal run release-readiness "This release has one change: https://github.com/autohealai/astronomy-shop-template/pull/1. Decide ship or hold." --commit <commit> --wait
See Tools and integrations and Model in the reference.