05 - Use Multiple Agents Safely#
Helmor supports running multiple agents in parallel — each in its own workspace. This multiplies throughput but requires discipline to avoid merge conflicts and review chaos.
The rule: one agent per workspace#
Never run two agents in the same workspace simultaneously. Helmor's concurrency lock prevents concurrent sends within a single workspace, but the real risk is logical — two agents editing overlapping files produce irreconcilable diffs.
Safe parallelism pattern#
- Separate workspaces — create one workspace per agent task, each branching from the same base.
- Non-overlapping scope — assign tasks that touch different files or modules.
- Sequential review — review and commit one workspace at a time. Don't batch-merge unreviewed work.
- Compare before merging — if two agents touched related code, compare their diffs before deciding which to keep.
Example: parallel feature development#
Workspace A (Claude Code) → "Add authentication middleware"
Workspace B (Codex) → "Add pagination to the list endpoint"
Both branch from main. Each agent works in isolation. You review A, commit, then review B and resolve any conflicts against the updated main.
When parallelism helps#
- Independent features with no shared files.
- Different parts of a large refactor (frontend vs. backend).
- Exploratory approaches — run the same task with two agents, compare results, keep the better one.
When to stay serial#
- Tasks that share state (same database migration, same config file).
- Small tasks where the overhead of a second workspace isn't worth it.
- When you're behind on review — more agent output without review is just unreviewed debt.
Key principle#
Parallelism helps throughput only when review stays clear.
If diffs are piling up faster than you can review, slow down and ship what's ready before starting more.
Related#
- Create a Workspace — set up isolated workspaces.
- Commit and Ship Work — the shipping step that gates parallelism.