Skip to content

Troubleshooting: Control 2.17 — Multi-Agent Orchestration Limits

Last Updated: April 2026

This guide covers the high-frequency failure modes observed when implementing the orchestration controls. The emphasis is on false-clean patterns — situations where the control appears to be working but is silently failing — because those are the highest-risk failures for FSI audit defensibility.

Quick Reference

Symptom Most likely cause Fix
Audit query returns zero events Wrong cloud endpoint, audit logging disabled, or record type renamed Verify -Endpoint / -ExchangeEnvironmentName; confirm Control 1.7 is implemented; cross-check current Microsoft Learn record types
Depth limit not enforcing Tracking variable not incremented or not passed to child agents Audit topic flow; confirm increment occurs before delegation
Circuit breaker stuck Open Reset timeout too long, downstream still failing, or half-open probe failing Manually inspect fsi_CircuitBreakerState; verify downstream health independently
Cascade failures still occurring Circuit breaker not invoked on every delegation edge Audit topic flow for missed fsi-CircuitBreaker-Check calls
HITL approvals abandoned by users Timeout too short, approver group unstaffed, or card lacks context Tune timeout; configure secondary approver chain; enrich card metadata
MCP tool invoked from unapproved server DLP gap or registry not synced to Copilot Studio Tighten Power Platform DLP (Control 2.14); audit registry sync
App Insights events missing dimensions Custom event call missing required custom-dimension keys Update topic / flow; re-test with Test 7 from Verification & Testing
Cross-environment depth tracking returns null Dataverse row not written or If-Match ETag conflict not handled Add retry-on-conflict in the flow; confirm row retention policy

False-Clean Patterns (Highest Audit Risk)

These patterns produce evidence that looks compliant but is not. Hunt them actively during quarterly reviews.

Pattern 1 — Sovereign-cloud endpoint omission

Symptom: PowerShell evidence script reports zero orchestration events; tenant clearly has Copilot Studio activity.

Root cause: Cmdlet authenticated against commercial endpoints from a GCC / GCC High / DoD operator workstation. Output is real but from the wrong tenant — likely empty or from an unrelated commercial tenant.

Detection: Compare the tenant ID returned by Get-MgContext (or Get-AdminPowerAppEnvironment | Select -First 1) against the expected target tenant ID before treating evidence as authoritative.

Fix: Always pass -Endpoint (Power Apps) and -ExchangeEnvironmentName (EXO) explicitly. See PowerShell baseline §3.

Pattern 2 — Depth variable not threaded through child agents

Symptom: Each individual agent appears to enforce depth correctly; Test 1 (depth violation) passes locally per agent. But end-to-end test exceeds the limit.

Root cause: Each child agent reinitializes OrchestrationDepth = 0 because the parent did not pass the current depth as an input parameter on invocation.

Fix: Confirm every delegation node passes Topic.OrchestrationDepth as an input. For Pattern B (cross-environment), confirm every agent reads from the same Dataverse row keyed by correlationId, not a per-agent local copy.

Pattern 3 — HITL timestamp drift indicating rubber-stamping

Symptom: HITL telemetry shows uniformly short waitDurationMs (e.g., consistently <30 seconds across all approvers).

Root cause: Approvers are clicking through without reviewing context. Audit-defensibility under FINRA Rule 3110 requires meaningful supervisory review, not just a click.

Fix: Sample 5% of approvals quarterly for quality review; consider adding mandatory free-text justification on the adaptive card; train approvers on context expectations.

Pattern 4 — Audit log retention shorter than 6 years

Symptom: Quarterly evidence pack succeeds for the 90-day window but historical reconstruction fails for prior periods.

Root cause: Microsoft 365 Audit (Standard) retains 180 days; Audit (Premium) retains 1 year by default. SEC 17a-4 recordkeeping for FSI typically requires 6 years.

Fix: Configure long-term retention per Control 1.7. Forward audit events to immutable storage (Sentinel + Storage with immutability policy, or Purview Data Lifecycle Management retention lock) to span the full 6 years independently of the M365 retention setting.

Pattern 5 — Telemetry emitted but not correlated

Symptom: Custom events visible in App Insights and audit records visible in Purview, but no way to reconstruct an end-to-end chain.

Root cause: correlationId not threaded consistently across (a) the originating agent, (b) Power Automate flows, (c) custom event emissions, and (d) Microsoft-emitted audit records (which use their own OperationId).

Fix: Standardize on a single correlationId field at the root agent and propagate through every downstream call. Where Microsoft-emitted IDs differ, persist the mapping (fsi_correlationId ↔ ms_operationId) in Dataverse so audit reconstruction can join across sources.

Detailed Issues

Issue: Delegation depth limit not enforcing

Diagnostic:

  1. Open the root agent topic in Copilot Studio.
  2. Confirm Topic.OrchestrationDepth exists as a topic-scoped variable initialized to 0.
  3. Confirm every delegation node is preceded by:
    • A Set variable node incrementing depth
    • A Condition node comparing Topic.OrchestrationDepth against the per-zone max
  4. Confirm the Condition node's true branch blocks the delegation (returns an error or routes to escalation), not merely logs and proceeds.

Resolution: If any step is missing, restore per Portal Walkthrough Step 3. For chains spanning environments, escalate to Pattern B (Dataverse correlation-ID-scoped depth).

Issue: Circuit breaker stuck Open

Diagnostic:

  1. Inspect fsi_CircuitBreakerState row for the affected targetAgentId. Note state, openedUtc, consecutiveFailures.
  2. Independently invoke the downstream agent (bypassing the orchestrator) and confirm health.
  3. If downstream is healthy: confirm the half-open probe flow is firing. The probe call must update state = 'Closed' on success.
  4. Check for ETag conflicts: if the probe and a fresh failure race, the probe's update may be lost.

Resolution: Add explicit retry-on-conflict to the probe flow. If timeout intervals are too long for operations, tune downward (but do not below 30s for Zone 3, otherwise probes can mask intermittent downstream issues).

Issue: HITL escalations missed by approvers

Diagnostic:

  1. Confirm the secondary approver group is populated and members have access to the channel/mailbox where adaptive cards are delivered.
  2. Verify the escalation path triggers automatically on timeout — not via a manual workflow step that may not be running.
  3. Sample escalation events in App Insights to confirm Orchestration.HitlDecision events fire with decision = 'EscalatedTimeout'.

Resolution: Treat any silently abandoned approval as a Control 3.4 incident (failure to obtain human approval before a customer-impacting action). Tune timeouts to realistic approval SLAs; never set timeout to "infinite" because that masks abandonment as in-progress.

Issue: MCP tool invoked from unapproved server

Diagnostic:

  1. Query App Insights for Orchestration.MCP.ToolInvocation events; group by mcpServer.
  2. Diff against the approved MCP server registry.
  3. If unapproved entries exist, identify which agent registered the tool and when (Copilot Studio agent activity log).

Resolution: Block at two layers:

  • Power Platform DLP (Control 2.14): place the MCP HTTP connector in the Blocked group for non-business-data environments; in the Business group for governed environments where it is needed.
  • Registry sync: if you maintain the MCP registry as a SharePoint list or Dataverse table, build a daily Power Automate flow that reconciles registered MCP servers in production agents against the approved list and opens an incident on drift.

Treat any drift event as a Control 3.4 incident and document in the quarterly attestation.

Known Limitations (April 2026)

Limitation Impact Mitigation
No native depth-tracking primitive in Copilot Studio All depth enforcement is custom (topic variables or Dataverse) Use Patterns A/B; treat the absence as control debt and lobby Microsoft via the Power Platform Ideas portal
No native circuit breaker Custom Power Automate flow required Documented in Portal Walkthrough Step 4
Limited cross-environment orchestration visibility Reconstructing chains spanning environments requires manual correlation Use Pattern B + correlation-ID propagation; consider keeping Zone 3 chains within a single environment
HITL via adaptive cards uses polling Latency between approval and resume Use Microsoft Agent Framework HITL primitives where available — they integrate with checkpoint persistence
Custom MCP servers in preview (April 2026) Behavior may change before GA Do not promote custom MCP servers to Zone 3 production until GA confirmed on Microsoft Learn
128-tool-per-agent ceiling Large delegation trees can hit ceiling silently Run Section 4 of PowerShell Setup monthly to track headroom

Escalation Path

If the issue persists after diagnostic steps:

  1. Power Platform Admin — environment, DLP, connector, or PPAC configuration questions
  2. Copilot Studio Agent Author — topic, variable, and tool-registration issues
  3. AI Governance Lead — policy, limits, registry, or attestation-quality questions
  4. Compliance Officer — supervisory-review (FINRA 3110), recordkeeping (FINRA 4511 / SEC 17a-4), or model-risk (OCC 2011-12 / Fed SR 11-7) interpretation
  5. Microsoft Support — platform behavior gaps, audit record-type changes, or suspected platform regressions; cite Microsoft Learn references and your tenant ID

For incidents involving customer impact, regulated data, or material financial decisions: open a Control 3.4 incident first, then troubleshoot.

Back to Control 2.17 | Portal Walkthrough | PowerShell Setup | Verification & Testing