Multi-Agent Mastery Path

Who it's for

Tech leads and senior ICs running more than one agent (or agent session) on the same initiative — feature factories, parallel refactors, or research+implement splits.

Prerequisites

• Fluency with single-agent spec → implement → verify (complete Senior Engineer first if not)
• A feature large enough that parallelism could help (not a one-file fix)
• Human willing to own integration and merge order

Ordered modules

1. Agent Roles — name responsibilities; don't clone the same prompt thrice
2. Coordination — interfaces, sync points, conflict rules
3. Multi-Agent Patterns — proven splits worth copying
4. Large Tasks — when to split at all
5. Task breakdown template — agent-sized units
6. Multi-agent task split — non-overlapping ownership
7. Context Reset note — kill thrash fast
8. Stuck Loops and Thrash — parallel thrash is expensive
9. Multi-Agent Feature case study — end-to-end story
10. Recovery — integrate or abort

Exercises

1. Take one epic; produce a task breakdown with explicit parallel vs serial edges and path ownership.
2. Run two agents on non-overlapping packages with a frozen interface file; integrate yourself; note every conflict cause.
3. Practice a mid-flight context reset on one thrashing agent using the reset note; compare token/time cost to continuing.
4. Write a post-integration checklist for your team (merge order, contract tests, final verify commands).

Capstone

Deliver one multi-agent feature to main: shared spec, frozen contracts, ≥2 agents with disjoint ownership, human integration, green verify, and a short retro (what leaked across boundaries, what you'd freeze earlier next time).

Expected outcome

You can parallelize agent work without merge chaos — ownership is explicit, interfaces are frozen, and recovery/abort criteria are decided before the second agent starts.

Related

• Learning Paths
• Senior Engineer path — prerequisite skills
• Scope and Context Failures
• Agent Handoff note