PowerLeap Logo

Copilot Studio Friction

Get notified when this record changes

One email when the status or the fix changes — double opt-in, no tracking, unsubscribe in every email.

Agent stops answering or loses citations after you change its instructions

Mitigatedsince 8 July 2026

Last verified

Details & related

Assessment

Confidence
Confirmed officialConfirmed by Microsoft documentation or an official statement.
Severity
Blocking
Typical time lost
Hours

Identification

OrchestrationAll channels

Verification & changes

  1. Verified

    Doc check by human: Generative mode guidance, orchestration FAQ and quotas page re-read during seed migration; the instructions-as-code warning and bisection advice are still documented.

  2. Change

    Provisionally approved by the Product Owner; external LLM quality review pending.

  3. Change

    Initial record created from the seed dossier (migration wave 1).

Related

Are you in the right place?

  • You edited the agent's instructions, and behavior broke right afterwards.
  • The agent stops answering some or all questions.
  • Or answers arrive without the citations that worked before.
  • Undoing your edit restores the old behavior.
  • There is no error code — the agent behaves differently, without complaint.

If instead answers vanished without any instructions change → see Agent gives no answer and no error message. If instead the agent invents facts → see Agent makes up answers. If instead only one channel shows old behavior after your change → see Teams answers with an old version. If instead the agent broke after a solution import → see Imported agent answers nothing.

What's happening

Microsoft's guidance is unusually blunt: the system treats agent instructions like code (documented as of 2026-07-08). One wrong sentence can break the machinery — especially citations and the decision whether an answer is shown. Instructions are not polite suggestions; they steer control flow inside the orchestrator. A single ambiguous rule can silence whole answer paths. Picture one wrong line in a recipe — salt instead of sugar — ruining the entire cake. The official debugging advice mirrors code practice: remove all instructions, then re-add them one at a time, testing after each. One more documented limit: custom instructions cannot override the default fallback message.

For technicians

The guidance names two concrete break surfaces: citations and answer suppression. Instructions are capped at 8,000 characters (documented as of 2026-07-08). Output is non-deterministic, so a broken instruction can pass one test and fail the next. Test every re-added block several times before you trust it.

How to fix it

Solution 1

Official workaround

Bisect your instructions

  1. Copy the full current instructions into a file outside Copilot Studio.
  2. Save that file with today's date — it is your rollback copy.
  3. Remove all instructions from the agent.
  4. Test the questions that failed.

✅ You should now see: the agent answers again, without your custom behavior.

  1. Re-add the first instruction block.
  2. Test again, several times — answers vary between runs.
  3. Repeat re-adding and testing, one block at a time, until the failure returns.
  4. Rewrite the block that triggered the failure; shorter and more literal wins.
  5. Test the rewritten block again before moving on.

✅ You should now see: all instruction blocks restored, with answers and citations intact.

Solution 2

Community workaround

Roll back to the last working version

  1. Restore the previous instructions text from your saved copy or documentation.
  2. Test the previously failing questions.
  3. Reintroduce your intended change later, in small steps, using the bisection above.

Check that it worked

Ask five questions that cover your agent's core jobs, including two that produce citations. Run them in the published channel, not only in the test pane. Expected: every question gets an answer, and citations render where they did before.

If it didn't work

  • The channel still serves the old version. Instruction changes need a publish plus propagation time. See Teams answers with an old version — propagation delay is suspect number one.
  • Randomness fooled your test. Non-deterministic output means one good answer proves little. Repeat each test question several times.
  • Two instructions conflict. Each block passes alone, but together they contradict each other. Re-add blocks in pairs to find the clash.
  • The answer is filtered, not broken. Answers can vanish silently without any citation issue. Check Agent gives no answer and no error message.

Prevent it next time

  • Version your instructions outside the platform: every change becomes a dated copy you can roll back to.
  • Change instructions in small steps and test immediately after each step.
  • Treat instruction edits like code review: a second pair of eyes before every publish.

Evidence

  • Official documentationlearn.microsoft.com

    Official guidance says the system treats instructions like code, warns of broken citations and suppressed answers, and recommends re-adding instructions one by one.

  • Official documentationlearn.microsoft.com

    The orchestration FAQ confirms agent behavior is driven by instructions and metadata, so wording changes directly change behavior.

  • Official documentationlearn.microsoft.com

    The quotas page caps agent instructions at 8,000 characters.