Agent Workbench — local operating layer for coding agents
Coordinate coding agents without adding another control plane.
Agent Workbench is being prepared as a local-first operating layer for real coding-agent work: tasks, route decisions, claim leases, prompts, handoffs, reports, review, and reusable completion memory live with the repository.
The problem it is designed to solve
Coding agents can produce code quickly, but multi-agent work still needs an operating layer. Maintainers repeatedly need to answer what is ready, who owns it, what context should be passed, what changed, what needs review, and what the next agent should remember.
Agent Workbench turns that coordination into inspectable repo-local artifacts instead of relying only on chat history or a hosted black box.
Planned product surface
Task memory
Markdown tasks, acceptance criteria, durable completion learnings, decisions, and review evidence that survive beyond one agent session.
Safe parallel work
Route decisions and claim leases make ownership visible before several agents start solving the same task.
Better handoffs
Generate target-specific prompts and handoff packets from the task instead of reconstructing context manually.
Local Control + Desktop
Inspect projects, setup changes, and operational state through local UI surfaces while keeping the repository as the source of truth.
See the product direction
Public release status
Agent Workbench is planned for public OSS release. The implementation is still being prepared behind a private repository while public packaging, documentation, licensing, and release gates are finalized. Public repository and installation links will appear here only after they are actually ready.
Product boundary
Agent Workbench is not intended to be another autonomous coding agent. It is the local operating layer around agent work so humans and other agents can inspect, resume, review, and improve the workflow.