feat: asign multiple coding agents the same task (to compete for my admiration) #46

New issue

Open

opened 2025-10-14 15:39:31 -06:00 by navan · 0 comments

navan commented

2025-10-14 15:39:31 -06:00

Owner

Originally created by @rawwerks on 10/2/2025

"Let me spin up multiple parallel attempts of the same task (one per agent/worktree) and then I pick and merge/combine the ones I like."

Motivation

Sometimes I just want quick side-by-side alternatives. Fastest path: create isolated attempts, let different agents hack, open PRs, and I merge the ones I like after a quick diff. This stays simple, keeps me in control, and avoids extra orchestration.

Proposed Solution

High-level overview
- Add a lightweight multi-attempt flow that duplicates one task into N isolated contexts (tags/worktrees), launches one agent per context, and opens a PR per attempt. I compare diffs and merge what I want.
Technologies / integrations
- Git worktrees + branch-per-attempt, Task Master tags, existing agent CLIs (Claude/Gemini/etc.).
Fit in current workflow
- Thin wrapper around what already exists:
  - Orchestrator creates N contexts/branches
  - Executors run one agent per context
  - Output is a PR per attempt; I review/merge

High-Level Workflow

Create attempts — Duplicate task <id> into N tags and optional worktrees; create matching git branches.
Run agents — Start one agent per context; stream logs as usual.
Open PRs — When an attempt hits “review,” auto-open a PR named after the agent/context.
Review & merge — I inspect diffs locally or in my VCS UI and merge any subset.
Cleanup — Optionally delete extra branches/tags/worktrees.

Key Elements

New minimal CLI
- multi-attempt start — create N contexts
- multi-attempt status — show per-attempt progress
- multi-attempt open-pr — one PR per attempt
- multi-attempt cleanup — remove extras
Lightweight data
- Task annotation: attempts: [{ tag, branch, agent, status }]
No judging, no scores
- Keep the UX to PRs + diffs + my choice.
Safe isolation
- Deterministic names: attempt-t<id>-<agent>
- Separate env vars/cache per attempt

Example Workflow

# 1) Create 3 attempts for task 42 with worktrees/branches
$ task-master multi-attempt start --task 42 \
    --agents "claude,gpt-gemini,qwen" \
    --create-worktrees

→ attempts: attempt-t42-claude, attempt-t42-gpt-gemini, attempt-t42-qwen
→ branches: feat/t42-claude, feat/t42-gpt-gemini, feat/t42-qwen
→ worktrees: ../t42-claude ../t42-gpt-gemini ../t42-qwen
→ executors launched (3)

# 2) Check progress
$ task-master multi-attempt status --task 42
→ claude: In Progress (3 commits)
→ gpt-gemini: Review-ready (tests pass)
→ qwen: In Progress (lint clean)

# 3) Open PRs (or auto on “review-ready”)
$ task-master multi-attempt open-pr --task 42
→ PR #121 feat/t42-gpt-gemini
→ PR #122 feat/t42-claude (draft)
→ PR #123 feat/t42-qwen (draft)

# 4) I merge the one(s) I like
$ gh pr merge 121 --squash --delete-branch

# 5) Cleanup leftovers
$ task-master multi-attempt cleanup --task 42 --keep "gpt-gemini"
→ removed tags/branches/worktrees for claude, qwen

*Originally created by @rawwerks on 10/2/2025* > "Let me spin up multiple parallel attempts of the *same* task (one per agent/worktree) and then **I** pick and merge/combine the ones I like." ### Motivation Sometimes I just want quick side-by-side alternatives. Fastest path: create isolated attempts, let different agents hack, open PRs, and I merge the ones I like after a quick diff. This stays simple, keeps me in control, and avoids extra orchestration. ### Proposed Solution - **High-level overview** - Add a lightweight **multi-attempt** flow that duplicates one task into N isolated contexts (tags/worktrees), launches one agent per context, and opens a PR per attempt. I compare diffs and merge what I want. - **Technologies / integrations** - Git **worktrees** + branch-per-attempt, Task Master **tags**, existing agent CLIs (Claude/Gemini/etc.). - **Fit in current workflow** - Thin wrapper around what already exists: - Orchestrator creates N contexts/branches - Executors run one agent per context - Output is a PR per attempt; I review/merge ### High-Level Workflow 1. **Create attempts** — Duplicate task `<id>` into N tags and optional worktrees; create matching git branches. 2. **Run agents** — Start one agent per context; stream logs as usual. 3. **Open PRs** — When an attempt hits “review,” auto-open a PR named after the agent/context. 4. **Review & merge** — I inspect diffs locally or in my VCS UI and merge any subset. 5. **Cleanup** — Optionally delete extra branches/tags/worktrees. ### Key Elements - **New minimal CLI** - `multi-attempt start` — create N contexts - `multi-attempt status` — show per-attempt progress - `multi-attempt open-pr` — one PR per attempt - `multi-attempt cleanup` — remove extras - **Lightweight data** - Task annotation: `attempts: [{ tag, branch, agent, status }]` - **No judging, no scores** - Keep the UX to PRs + diffs + my choice. - **Safe isolation** - Deterministic names: `attempt-t<id>-<agent>` - Separate env vars/cache per attempt ### Example Workflow ```shell # 1) Create 3 attempts for task 42 with worktrees/branches $ task-master multi-attempt start --task 42 \ --agents "claude,gpt-gemini,qwen" \ --create-worktrees → attempts: attempt-t42-claude, attempt-t42-gpt-gemini, attempt-t42-qwen → branches: feat/t42-claude, feat/t42-gpt-gemini, feat/t42-qwen → worktrees: ../t42-claude ../t42-gpt-gemini ../t42-qwen → executors launched (3) # 2) Check progress $ task-master multi-attempt status --task 42 → claude: In Progress (3 commits) → gpt-gemini: Review-ready (tests pass) → qwen: In Progress (lint clean) # 3) Open PRs (or auto on “review-ready”) $ task-master multi-attempt open-pr --task 42 → PR #121 feat/t42-gpt-gemini → PR #122 feat/t42-claude (draft) → PR #123 feat/t42-qwen (draft) # 4) I merge the one(s) I like $ gh pr merge 121 --squash --delete-branch # 5) Cleanup leftovers $ task-master multi-attempt cleanup --task 42 --keep "gpt-gemini" → removed tags/branches/worktrees for claude, qwen