JKP’s multi-agent mode is now my default for big refactors. Multi-agent mode: 3 AI agents collaborate on big code refactors. Here is the workflow, the results, the time savings.
What is multi-agent mode
JKP lets me spawn 3 AI agents that work on the same refactor in parallel. Each agent has a different role:
- Agent 1 — the architect. Designs the refactor. Identifies the abstractions. Proposes the new file structure.
- Agent 2 — the implementer. Writes the code based on Agent 1’s design. Runs the test suite.
- Agent 3 — the reviewer. Reviews Agent 2’s code. Suggests improvements. Catches edge cases.
The workflow
- I prompt JKP with the refactor. Example: “Refactor the EFURC card renderer to support the new Exile mechanic.”
- JKP spawns 3 agents. Each agent works in parallel on its role.
- Agent 1 produces a design doc. Agent 2 produces the code. Agent 3 produces the review.
- JKP merges the 3 outputs. Final result: design doc + code + review.
Time savings
Single-agent mode: 4 hours for a big refactor (1000+ line code change).
Multi-agent mode: 1 hour. 4x speedup.
The speedup comes from parallelism. Agent 1 designs while Agent 2 waits. Agent 2 implements while Agent 3 reviews. The agents don’t block each other.
Quality comparison
Single-agent mode: 87% pass rate on the first test run.
Multi-agent mode: 96% pass rate on the first test run.
Agent 3’s review catches edge cases that Agent 2 misses. The review pass is the difference between 87% and 96%.
What this means for AI-assisted game dev
Multi-agent mode is the next step in AI-assisted game dev. Single-agent mode is fast. Multi-agent mode is fast AND accurate. The combination is what makes the AI agent viable for production code.
What this means for EFURC v0.3 (Outsiders)
The EFURC v0.3 Outsider refactor (new cards + the Exile mechanic) is being built with multi-agent mode — one design pass, one implementation, one review, in parallel.
Read what Jester King Prime actually does.
