Taskplane is an open-source multi-agent coding orchestration system built on Anthropic's Pi agent SDK. It transforms a batch of tasks into dependency-aware waves of parallel work, each lane running in an isolated Git worktree, with cross-model review gates before automated merges.
Why It's in Assess
Taskplane introduces a distinctive "light-factory" architecture that solves two hard problems in long-running AI coding tasks: context exhaustion and merge chaos. Its use of Git worktrees for full filesystem isolation, combined with PROMPT.md/STATUS.md files as a persistent memory store, means workers survive context resets without losing state. A four-role agent topology (Supervisor, Workers, Reviewers, Merger) with intentional cross-model review (different model for reviewer than worker) adds a quality gate that most single-agent tools lack.
Team model vs. fleet model: Most parallel-agent tools (Composio agent-orchestrator, Superset) use a fleet model — one independent agent per issue, each with its own branch and PR. Taskplane's four-role topology is a team model: multiple agents with specialized roles collaborate on the same body of work. The Supervisor monitors the whole batch and can intervene, Workers execute individual tasks, Reviewers independently audit each Worker's output before merge, and the Merger handles integration. This makes Taskplane better suited to tasks that benefit from oversight and review within a single coherent project, rather than fully independent parallel issues.
The system is in active development (133 releases, v0.28.2, 1,488+ commits) but has a small community footprint (107 GitHub stars, 2 forks as of April 2026) — it's early-stage and Pi-SDK-dependent, which limits portability. Worth evaluating for complex, multi-repo coding workflows where vanilla Claude Code or a single-agent loop breaks down.
Key Architecture Concepts
- Dependency DAG → waves → lanes: Tasks parsed into a dependency graph, serialized into waves, then run in parallel lanes within each wave. Mono- and polyrepo support, with per-repo worktree isolation for polyrepo batches
- Persistent file-based memory:
PROMPT.md(mission, steps, constraints) andSTATUS.md(progress tracking) survive context resets — workers re-read them on restart, preserving continuity across long tasks - Four-role topology: Supervisor monitors progress and handles failures autonomously; Workers execute tasks; Reviewers provide inline feedback at step boundaries (cross-model by design); Merger resolves conflicts and verifies integration
- Step-boundary commits: Prevents work loss; each completed step is committed before moving to the next
- Live web dashboard: Real-time SSE-streamed monitoring at
localhost:8099— lane/task progress, reviewer activity, merge telemetry, batch history
Caveats
- Requires Pi (Anthropic's agent SDK) — not framework-agnostic; tight dependency limits portability to other agent runtimes
- Node.js 22+ and Git required; setup is developer-oriented, not plug-and-play
- Community is tiny (107 stars) and the project is solo-authored; long-term maintenance is uncertain
- No official pricing/enterprise tier — purely open-source BYOK today
Key Characteristics
| Property | Value |
|---|---|
| Interface | CLI + web dashboard |
| License | MIT |
| Underlying model | Any (via Pi SDK) |
| Isolation | Git worktrees (per lane) |
| Pricing | Free (BYOK) |
| Provider | HenryLach (individual) |
| GitHub | HenryLach/taskplane |
| npm | taskplane |