Playbooks · Coding
Triage issues with a read-only agent
Written October 31, 2026
A small team with a pile of issues waiting for triage, and a Copilot agent that can only read. The agent checks the evidence in each issue before anyone applies a label — and because it has no write tools, it cannot edit, label, comment, or open a pull request. Triage is a reading job. Give it to a reader.
Why triage needs a read-only agent
The obvious way to hand an issue to Copilot is the wrong one for triage. Assigning an issue to Copilot always creates a pull request. Triage does not need a pull request. It needs an answer to a narrower question: does this issue carry the evidence for the label it asks for? A default agent, given a pile of issues, can act on them. A read-only agent can only report. That is the whole point. The human keeps the only write step in the process: applying the label, after checking the evidence the agent cites.
Set up the read-only agent
Build the agent as a custom agent: a Markdown profile with YAML frontmatter that names the agent, describes it, and lists the tools it may use. The tools list is the restriction, so write it narrowly:
- Go to the agents tab, select the repository, and create an agent. A template profile called
my-agent.agent.mdopens in the repository's.github/agentsdirectory. - Name the file
triage-reader.agent.md. File names for agent profiles may only contain letters, numbers, dots, hyphens, and underscores. - Set the frontmatter. The description is required. The tools list must contain only reading tools:
--- name: triage-reader description: Reads issues and checks the evidence for a label, without changing anything tools: ["read", "search"] --- You are a triage reader. Read issue text and repository code. Never edit a file, never apply a label, never comment on an issue, and never open a pull request. Report evidence only: quote the issue text you relied on, name the issue numbers, and point at the code or check output. Where the evidence is missing, say so plainly.
If the profile omits the tools property, the agent gets access to all available tools — so the list above is load-bearing, not decoration. Commit the file and merge it into the repository's default branch, then go back to the agents tab and refresh the page: the agent appears in the dropdown in the prompt box, and can be selected there instead of the default agent.
Run a triage pass
Start from a prompt in the agents tab or agents panel. Do not assign the issues to Copilot: assignment always creates a pull request. Select the repository, select the triage-reader agent, and list the issues to check, by number, with the question to answer for each. The agent's report must carry, for every issue:
- The proposed label, if any — one of the repository's labels, for example the bug, duplicate, or question label.
- The exact evidence — a quote from the issue text, the issue number, and, where the claim touches code, the file and the lines, or the output of a check.
- What is still missing — the reproduction steps, the version number, or the log that would settle it.
- A plain "cannot tell" verdict where the evidence is absent. A missing verdict is how an unproven label gets applied.
Two rules keep the report honest. A duplicate verdict needs the matching issue number and overlapping evidence, not a similar title. And a report that says "bug" with no reproduction steps gets the question verdict, not the bug label: the question label means the issue needs more information, which is exactly the state it is in.
Apply the labels yourself
Read the report, open the cited evidence, and apply the labels. Anyone with triage access can apply and dismiss labels from the issue sidebar. From the command line, gh issue edit with --add-label adds labels, but the labels must already exist in the repository, and multiple labels are separated with commas. If the report proposes a label the repository does not have, create it first — anyone with write access can create a label — then apply it. The agent never applies a label. If its report is wrong, the cost is a wrong paragraph. If it could label, the cost would be a wrong label on a live issue.
A worked example: three issues, one pass
Issue 41 reports a crash on startup, with a stack trace. The agent quotes the trace, finds the same trace quoted in issue 38, and proposes duplicate with 38 as the match. A human opens both, confirms the traces match, and labels 41.
Issue 42 reports that saving fails, with no steps, no version, and no error text. The agent proposes no label, lists what is missing, and records "cannot tell". A human labels it question and asks the reporter for the missing pieces.
Issue 43 is titled as a bug, but its body is a request for the behaviour of an option, with no failing case. The agent quotes the body, notes there is no reproduction, and proposes question. The title said bug. The evidence said question. The evidence wins.
What can go wrong
- Assigning the issues to Copilot instead of prompting. Assignment always creates a pull request, and triage ends up with code changes nobody asked for.
- Omitting the tools list. A profile with no tools property gets all tools. The read-only property lives in that list.
- Letting the prompt soften the restriction. The profile's instructions forbid edits, labels, comments, and pull requests. A prompt that asks the agent to "go ahead and fix it" asks it to break its own profile. Keep the prompt to reading and reporting.
- Applying a label from the report without opening the evidence. The report is a pointer to evidence, not the evidence. The human check is the step that makes the label trustworthy.
- Accepting a duplicate verdict on a similar title. Duplicates need the matching number and overlapping evidence. Similar wording is not a match.
Limits
- The agent reads text and code. It does not run the program, so reading the code is not the same as reproducing the bug. A code path that looks wrong is still unproven until someone reproduces it.
- The agent works from the issue text as given in the session. If an issue is edited after the session starts, ask again rather than assuming the agent saw the edit.
- Evidence from code is not a reproduction, and a quote from an issue is not proof the report is accurate. Both are pointers a human checks before labelling.
- Copilot cloud agent is available for paid Copilot plans, and in all repositories except those owned by managed user accounts and those where it has been explicitly disabled. A team outside those bounds cannot run this playbook as written.
Related playbooks
- Triage a backlog of issues with an agent
- Thousands of agents read a mouse-brain atlas for leads
- Gemini Robotics 2.5: 96% packing, partners only
- Haiku 5.5 is 90% cheaper only under 100K tokens: price the tier jump and the tokenizer before you route background work to it
Sources
Sources: Kick off a task with Copilot agents on GitHub, Creating custom agents for Copilot cloud agent, Your first custom agent, Managing labels, Adding labels to issues.
Back to: Coding playbooks hub