Managing tasks an AI agent discovers mid-work

Managing tasks an AI agent discovers mid-work comes down to one rule: record the discovery as a new, linked task and keep it out of the current one. Agents find things constantly. While adding a VAT line to an invoice, an agent notices that the CSV export rounds money with a different helper, that a test fixture is out of date and that a function name is misleading. Left alone, it will often fix some of them on the spot. The current task then grows without anyone deciding it should, the diff mixes unrelated changes and the review gets harder. The opposite failure is just as common: the agent mentions the problem in its chat output, the session ends and the finding is lost. The four steps below keep every finding without letting it change what the current task is allowed to do. They work with any agent and any task tracker. For keeping the resulting backlog tidy, see backlog hygiene for AI projects.

Step 1: Ask the agent to report, not fix

Put an explicit instruction in the task or the agent's standing instructions: 'If you find work outside this task, do not do it. List it under Discovered in your report.' Then ask for a fixed shape for each item, so findings are comparable and easy to turn into tasks.

A useful shape has five parts. What: 'CSV export uses its own rounding instead of cents()'. Where: 'src/export/csv_rows.py, line 41'. Why it matters: 'export totals can differ from checkout totals by one cent'. Blocking: whether this task cannot be honestly completed without it, yes or no. Evidence: the test, log line or code that shows it.

The instruction does not need to be clever. Its job is to give the agent a legitimate place to put its observation, so fixing it in place stops looking like the only helpful option. Pair it with a declared scope, as described in keeping agents within scope, so a fix that slips through shows up as an out-of-scope path in the diff.

Step 2: Classify each finding as required or optional

Required means the current task cannot be completed honestly without it. If the VAT line depends on a rounding bug being fixed first, the finding is a prerequisite. Record it as a dependency: the current task now depends on the new one and should not be marked complete until the new task is done, or it should be split so the part that does not need the fix can finish.

Optional means the finding is worth doing but the current task is complete without it. The misleading function name is optional. Record it as a related link, so someone reading either task later can find the other, with no effect on whether the current task can complete.

The classification is a judgment, and agents get it wrong in both directions. They mark cosmetic issues as blocking, and they miss real prerequisites because they did not run the check that would have shown them. Review the classification when you review the task. A wrong 'required' blocks work for no reason; a wrong 'optional' lets a task complete on top of a known problem, which is how a fix breaks another feature.

Step 3: Look for an existing task before creating one

Before creating a new task, search for one that already covers the finding. Search by file path first, because paths are stable: 'csv_rows.py' finds earlier tasks about the export far more reliably than prose does. Then search by the key nouns: 'rounding', 'cents', 'export'.

Expect the search to miss things. Two agents describe the same problem in different words ('unify money rounding' versus 'export totals off by a cent'), and keyword search will not connect them. A quick human look at the candidate list is worth more than an extra automated pass. If an equivalent task exists, link the finding to it and add the new evidence there instead of creating a duplicate.

Duplicates are expensive with agents in particular. Two tasks for the same fix can both be picked up, producing two competing changes in the same file. That is one of the common causes behind an agent repeating work that was already done.

Step 4: Create the task without permission to run

A new task created from a discovery should start in a planning state with no execution scope. It is a proposal. It needs its own scope and its own approval before any agent works on it, exactly like a task you wrote yourself. This matters most for unattended work: if a runner can pick up any ready task, a discovery becomes work that nobody approved. Keep discovered tasks off the list of tasks a run is authorized to take, as covered in limits for unattended runs.

Write the records in a safe order. Create the new task first, then add the dependency from the current task to it. If the process is interrupted between the two writes, the worst case is an unlinked new task, which is easy to spot. The reverse order can leave the current task depending on a task that was never written.

Finally, never widen the current task to absorb the discovery after the fact. If you decide the fix belongs in this change, edit the scope deliberately, approve it again and note why.

QUESTIONS

Should an AI agent fix bugs it finds outside its task?

Generally no. Ask it to report the bug with the file, the evidence and whether the current task depends on it, then create a separate task. Fixing in place mixes unrelated changes into one diff, skips scoping and approval for the fix, and makes the review harder. The exception is a fix the current task cannot work without, and that should be a deliberate scope change.

What is the difference between a required and an optional follow-up task?

A required follow-up is a prerequisite: the original task cannot be honestly completed until it is done, so it becomes a dependency. An optional follow-up is useful but independent: it gets a related link and has no effect on whether the original task can complete. Getting this wrong either blocks work needlessly or lets a task complete on a known problem.

How do I stop an agent from creating duplicate tasks?

Make it search before it creates. File paths are the most reliable search key, followed by the main nouns of the problem. Expect keyword matching to miss duplicates phrased differently, and have a person glance at likely candidates. When an equivalent task exists, link the finding and add the evidence there.

RELATED GUIDES

• How to create a project card in BRAIN FLUID Desk
• How to find and organize work across projects

RELATED HARNESS PAGES

• Keeping an AI project backlog clean Keeping an AI project backlog clean: decline or hold work instead of deleting it, merge duplicates, avoid ID collisions and groom on a schedule.
• Keeping AI coding agents within scope Keeping AI coding agents within scope: declare paths, approve them separately with an expiry, check writes during the run and compare the diff after.
• Why does my AI coding agent keep repeating work? Agents repeat work when nothing records what was done or declined, tasks are duplicated, or context is lost. How to tell which, and fix it.
• Why a parent task can't be done before its children Why a parent task can't be done before its children: required versus related work, and why finishing a child changes eligibility, not status.