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 suddenly stops staying in the topic without any change on your side

Regressedsince 8 July 2026

Last verified

Details & related

Assessment

Confidence
CorroboratedMultiple independent sources describe the same behaviour.
Severity
Degrading
Typical time lost
Days

Identification

OrchestrationAll channels

Verification & changes

  1. Verified

    Doc check by human: Bounded web check: the Q&A regression thread is still online with no fix announcement found; regression last confirmed Nov 2025; re-verification in own tenant pending.

  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?

  • Your agent held topics across follow-up questions for weeks or months.
  • It suddenly answers follow-ups with "I can't find a related topic" or similar.
  • Sometimes it starts generating the answer, aborts, and asks the user to rephrase.
  • You changed nothing: no new topics, no new instructions, no publish.
  • Community threads report the same change in the same period.

Ask yourself: since when? If the agent has always jumped when an answer matched another trigger phrase → see Topic exits mid-question. That behavior is permanent and by design. If instead routing broke right after "End all topics" → see Wrong topic after End all topics. If instead only one channel shows old behavior after a publish → see Teams answers with an old version.

What's happening

This record tracks a dated platform change, not a design flaw you built in. In November 2025, multiple makers reported the same break: agents suddenly stopped staying in the active topic. Follow-up questions landed nowhere. Some agents began generating an answer, aborted, and demanded a rephrase — behavior nobody had configured. Nobody changed their agents; the platform changed under them. Picture a train timetable the operator rewrites overnight without telling anyone. Copilot Studio ships monthly updates, and routing or context retention can drift with them. Microsoft Support was investigating the cause with affected makers as of November 2025. As of 2026-07-08 we found no fix announcement, so the regression stays on the books.

For technicians

The primary thread (Nov 2025) carries multiple confirmations, including a maker working with Microsoft Support on the root cause. An independent thread from the same period reports the same pattern.

Topic routing "was working correctly just a few days ago" and changed although nobody touched the agent. The suspected surface is orchestrator routing and conversation-context retention, not maker configuration. There is no error code; the failure is purely behavioral. Generative orchestration uses only the last 10 conversation turns, which caps any context-based self-help (documented as of 2026-07-08).

How to fix it

Solution 1

Community workaround

Confirm platform drift before touching your agent

  1. Write down the date and time the behavior changed.
  2. Check the Copilot Studio release notes for changes around that date.
  3. Search Microsoft Q&A for fresh threads describing the same symptom.
  4. Confirm with your team that nobody changed topics, instructions, or tools in that window.

✅ You should now see: evidence that the change came from the platform, not from your configuration.

Solution 2

Community workaround

Report it and anchor context while you wait

  1. Save the conversation ID of a failing conversation.
  2. Save the transcript of that conversation.
  3. Open a Microsoft support case with the ID, the transcript, and the date the behavior changed.
  4. Meanwhile, restate key context in each follow-up prompt so the agent can re-anchor.
  5. Where a jump hurts most, protect the node as described in Topic exits mid-question.

✅ You should now see: a dated, evidence-backed support case, and fewer failed follow-ups in daily use.

Check that it worked

Rerun the follow-up dialogue that used to work, in the published channel. Expected while the regression stands: it may still fail — this test tells you whether Microsoft's side changed. Retest after each platform update and note the date. Several clean days in a row support a status change to fixed.

If it didn't work

  • Wrong order. Check release notes and Microsoft Q&A before debugging your own configuration. Platform drift moves faster than your settings.
  • The channel serves an old version. Propagation delay is suspect number one for stale behavior. If you published anything, see Teams answers with an old version.
  • You face the by-design jump. If the agent always behaved like this, see Topic exits mid-question.
  • The workaround masks the signal. Restated context can hide the regression in tests; verify with plain, unaided follow-ups too.

Prevent it next time

  • Apply the drift rule to every sudden behavior change: first release notes and Q&A, then your own configuration.
  • Keep a dated log of unexplained behavior changes; it turns "the agent got weird" into a support-ready regression report.

Evidence

  • Community threadlearn.microsoft.com

    Multiple makers report (Nov 2025) that agents suddenly stop staying in the active topic on follow-ups, with no maker-side change.

  • Community threadlearn.microsoft.com

    An independent thread reports topic routing that 'was working correctly just a few days ago' changing without any maker action.

  • MVP blogjatinmadan.com

    A practitioner writeup (Mar 2026) documents that makers need extensive logging and telemetry to understand orchestrator routing changes.