GitHub Copilot Train-the-Trainer

Orchestrate agents and workstreams

Product and delivery teams, or developers after Session M2-05

Session M7-05 | 3 hours

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Agenda

Segment Time
Parent objective and boundaries 10 min
Minimal workstream plan 10 min
Monitor and intervene 12 min
Live demonstration 13 min
Review and lab handoff 15 min
Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Required access

  • /orchestrate must create and coordinate live child sessions.
  • Product route: GitHub MCP must read the selected issues.
  • Developer route: local project read access and Python 3.10 or later.
  • You must be able to open, redirect, and stop child sessions.

Stop if orchestration or the selected route's inputs are unavailable.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Choose one lab route

Route Inputs Result
Product and delivery Four live GitHub issues Parent result used by Session M7-06
Developer Local evaluator, tests, synthetic PR snapshot Separate implementation, coverage, and PR conclusions

Each route takes 120 minutes. Developers need Session M2-05, not Sessions M7-01 to M7-04.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

One parent objective

Review the four approved status-visibility issues and return one reviewed
result for each.

The parent owns scope and review.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Keep the plan small

Each workstream needs:

  • a question with a clear scope;
  • approved inputs;
  • a stop condition;
  • a reviewer.

The orchestration surface already holds session links, status, timing, and usage.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Four independent workstreams

Workstream Question
Public mapping Is the mapping ready for approval?
Requester status Is the issue ready for implementation review?
Empty and stale states Are fallback states testable?
Access boundary Are the approved requester-access results ready?

Move shared questions back to the parent.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Developer workstreams

Question Input
Does implementation meet the behavior contract? Brief and evaluator
What do tests prove or miss? Brief, tests, and evaluator
Is the supplied PR ready? Brief and PR snapshot

Children read the same fixed inputs. They do not wait for one another or edit.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Practise steering

Redirect the test reviewer after its first answer:

Check criterion 5. Does the test prove sorted reason codes, the exact complete
reason list, or both? Do not edit files.

Inspect the revised answer. A follow-up question does not prove the first answer
was wrong.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Reconcile the readiness claim

The synthetic author says: "Local tests pass, so the PR is READY."

Check the actual snapshot:

  • required integration-tests failed;
  • blocking thread R1 is unresolved.

Check the parent's decision against the snapshot. Do not manufacture conflicting
child answers. Local success does not override PR blockers.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Start live orchestration

/orchestrate Use the four workstreams in my approved plan. Keep them independent.
Each result must answer its question, cite approved inputs, state unresolved
gaps, and give a recommendation. Obey each stop condition.
Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Monitor on the platform

Signal Action
Safe and on scope Wait
Safe work needs correction Redirect
A boundary is crossed Stop
The result meets the plan Accept
The result lacks support Reject
Human authority is required Escalate

Do not maintain a second status log.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Review each result

  • Does it answer the assigned question?
  • Does it use only the approved issues and their linked sources?
  • Did it obey the stop condition?
  • Are gaps listed?
  • Is the recommendation supported?

Reject answers that lack supporting sources.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

The parent result is the handoff

The reviewed result:

  • links findings to live GitHub issues;
  • shows rejected claims;
  • names gaps and owners;
  • recommends the next follow-up.

Session M7-06 reads this result and the live issues.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

The human keeps authority

Copilot coordinates work. A person reviews results and approves any issue update,
publication, or release statement.

Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Session M5-01 is different

Session M5-01 Session M7-05
Persistent roles and routing Temporary workstreams
Optional framework setup Built-in orchestration
Shared memory Live issues and their supporting sources
Session M7-05: Orchestrate Agents and Workstreams
GitHub Copilot Train-the-Trainer

Lab deliverable

  1. Approved orchestration plan
  2. Live child-session results
  3. Reviewed parent result linked to live issues

Keep progress on the orchestration surface. Do not copy it into a delivery-status
form or copy session identifiers.

For developers, use the developer plan and reviewed parent result instead of
live issues. Include the actual redirection, rejected or unresolved claims, and
observed usage.

Session M7-05: Orchestrate Agents and Workstreams