Lesson 1 of 6 · 0%Decompose work into outcomes and dependency gatesNext
Course map

Multi-Agent Workflows, Goals and Production Operations

0 of 6 complete0 of 6

Lesson 1.1 · 45 minutes

Decompose work into outcomes and dependency gates

Replace a vague multi-agent request with accountable artifacts, explicit dependencies, and a human-owned release decision.

Skip course map
Current lessonDecompose work into outcomes and dependency gates0% complete · 0/6 lessons

Multi-Agent Workflows, Goals and Production Operations

0% complete · Current: Decompose work into outcomes and dependency gates

Verifiedon 2026.7.1-2

Action boundary

Before you act

Expected result
A graph names a deliverable owner, input, acceptance test, and stop condition for every node before any agent starts.
Failure mode
Agents are named before outputs are defined, so duplicate work and unreviewable summaries accumulate without a clear release decision.
Rollback
Stop unbounded tasks, preserve their artifacts as drafts, collapse duplicate nodes, and restart only after the graph has a named owner and acceptance test.

Start with the decision, not the cast

“Use several agents to watch releases” is not an operating design. Start with a decision that a named human will make: publish, hold, or discard a weekly release bulletin. Then work backwards through the artifacts required to make that decision defensible.

For the capstone, the minimum graph is:

  1. Source intake: a researcher produces a URL ledger of official release notes and documentation changes.
  2. Claim review: a verifier accepts, corrects, or rejects each proposed claim against the source URL and version.
  3. Draft: an editor turns accepted claims into a short bulletin with links and uncertainty labels.
  4. Publication decision: the publication owner approves, requests changes, or holds it. This node is human-only.

A node is ready only when its required predecessor artifacts exist and meet their stated test. A handoff saying “done” is not evidence; an artifact and test result are.

Lab: draw the dependency graph

Use one rectangle per artifact, not per agent. On each rectangle write: input fingerprint, accountable role, acceptance test, maximum elapsed time, and stop action. Draw solid arrows only where a later artifact cannot be judged without an earlier one. Add a red stop node for a missing primary source, a claim without a receipt, a budget breach, or any attempted publication without approval.

Check: Can the editor begin when source intake is only partially complete? If yes, state exactly which accepted claims are available and label the draft incomplete. If no, make the intake-to-review edge a hard dependency.

Checkpoint

Which definition is actionable?

  • “A research agent finds important news.”
  • “The researcher returns at most eight primary URLs, a one-sentence factual claim per URL, and a retrieval timestamp; entries without a source URL are rejected.”

The second definition wins because another role can verify it without guessing intent.

Learner artifact

Create the original dependency graph described in the course map. Put the human publication gate after the draft, never inside the editor node. Label any optional enrichment path as non-blocking so it cannot delay critical evidence.

Source provenanceVerification and sources

Review receipt rr_multi_agent_outcomes_graph

Outcome
approved
Method
source-review
Reviewer
scout-independent-review
Reviewed

Evidence

Limitations

  • Draft editorial review; no production workflow was connected or published.

Open the public evidence snapshot

Lesson checkpoint

Ready to move on?

Mark this lesson complete when you can apply its outcome without relying on the examples above.