AI agent workflow tools, mapped.
Every tool here answers one question differently: when the agent says “done”, what happens next? Most take its word. A few ask a person. Almost none run a check. That one question sorts the whole field — here it is, dated August 2026.
One question sorts the field
There are a lot of these tools now — GitHub’s spec-driven-development topic alone holds over 1,600 repositories as of August 2026. Lists of them sort by popularity or by which agent they support. More useful: ask what happens at the moment the agent claims the work is finished. Three answers exist.
Nothing. The claim is accepted. The spec, plan, or task list shaped the work going in, and nothing checks the work coming out.
A person approves. The tool stops and waits for a human to read the result and click approve. Honest, but the human is the check — and after the tenth approval, nobody is really reading.
A check runs. Tests re-run, evidence is demanded, or an independent reviewer inspects the claim — and completion is blocked until it passes.
The map
Figures from the GitHub API, 2026-08-17. Stars rounded; “quiet” means no push for three months or more at that date. Rows are grouped by the answer in column three — every “nothing” first, then the ones that gate — and by size inside each group, which is why the star counts are not one descending run. Tools below are the notable representatives, not the whole field. Where we have a page on a tool, its name links there and the ↗ goes to the repo.
| Tool | What it adds | When the agent says “done” | Status, August 2026 |
|---|---|---|---|
| superpowers ↗ | Skill library incl. process skills | Nothing — skills are advice | ≈273k stars, active |
| GitHub Spec Kit ↗ | Spec → plan → task documents | Nothing — documents are inputs | ≈130k stars, active |
| autoresearch ↗ | Autonomous research loop | Nothing | ≈94k stars, quiet since Mar |
| OpenSpec ↗ | Spec-driven change proposals | Nothing — review is convention | ≈65k stars, active |
| BMAD-METHOD ↗ | Agile role prompts + phase docs | Nothing — roles are prompts | ≈52k stars, active |
| claude-task-master ↗ | Task decomposition + tracking | Nothing — status is self-reported | ≈28k stars, quiet since Apr |
| gsd-core ↗ | Discuss-first loop; hooks guard its planning files | Nothing re-checks the claim | ≈8.4k stars, active |
| planning-with-files | Plan files + opt-in stop gate | Opt-in gate + attestation ledger | ≈26k stars, active |
| spec-workflow-mcp | Spec workflow over MCP | A person approves each document | ≈4.3k stars, active |
| tdd-guard | One job: no code before tests | A hook blocks the violating write | ≈2.3k stars, active |
| zeroshot | Executor / verifier separation | Isolated validators re-check claims | ≈1.7k stars, active |
| babysitter | Workflow as code, 12 harnesses | Quality gates before progression | ≈1.7k stars, active |
| team.management | Protocol engine on your process | Completion gate re-checks the diff; tests too, once configured | open source, active |
The empty end of the map
Here is the asymmetry the table understates. The spec end holds hundreds of thousands of stars across a dozen mature projects. The verification end — tools whose whole job is checking that claimed work happened — is close to empty: as ofAugust 2026, proof-loop has 6 stars, railwarden has 2, and agent-gates has 0. Nearly everyone tells the agent what to do. Almost nobody checks it did it.
Also worth knowing: this is not the only comparison. spec-compare maintains a side-by-side of 18 spec-driven tools with a live visualizer. It sorts by features rather than by the done-question, and it is good. Read both.
Where the engine sits
team.management is a protocol engine on top of the agent you already run. Your process becomes explicit steps; the engine holds the agent to them. At the completion step it does not take the agent’s word: the task protocol’s code-review gate re-runs verification — spec compliance against the task’s written criteria, completion evidence, and your test command once you have configured one — and refuses to advance while any of it fails. The mode gating behind it is DAIC.
The honest limit, stated plainly: enforcement binds the agent inside the run, on your machine — it is not portable proof, and the person who owns the config can always change it. What it removes is the silent failure, where a step was skipped and the summary said otherwise.
Most tools in the table compose with it rather than compete: a spec tool shapes what to build, the protocol engine controls how the work proceeds and checks that it happened.
FAQ
What is the difference between spec-driven tools and verification tools?
Spec-driven tools structure the work before it starts — specs, plans, task lists the agent is asked to follow. Verification tools check the work after the agent claims it is finished — re-running tests, demanding evidence, or blocking completion until an independent check passes. The first category is crowded (GitHub’s spec-driven-development topic holds over 1,600 repositories as of August 2026); the second is nearly empty.
Which AI agent workflow tools are still maintained in 2026?
Checked against the GitHub API on 2026-08-17: Spec Kit, OpenSpec, BMAD-METHOD, gsd-core, superpowers, tdd-guard, zeroshot, babysitter and spec-workflow-mcp were all pushed within recent weeks. claude-task-master (April), autoresearch (March), ccpm (March) and cc-sessions (October 2025) had gone quiet, and get-shit-done is archived — its successor is open-gsd/gsd-core. Check the repo before you adopt; this field moves monthly.
Do I have to choose one workflow tool?
Mostly no — the categories compose. A spec tool decides what to build, a process engine controls how the work proceeds, and a verification gate checks it actually happened. team.management sits in the second and third: protocol steps the agent moves through in order, with a completion gate that re-checks the work before it can be called done — an independent spec-compliance review of the diff against the written criteria, plus your test command if you have configured one.
Facts and figures on this page are as of August 2026, verified against the sources linked inline. If you’re reading this much later — check the sources.
Related comparisons
Deeper dives on tools this map places, and the two failures that sort the whole field.