Compare
codeloop next to the two kinds of agent-workflow tool it is most often weighed against. Where it is ahead, and where it is not.
Two kinds of tool are common today. Spec-first tools give an agent a sequence of slash commands that write a specification, a plan and a task list before any code, with one feature per branch and folder. Agile-persona tools give an agent a cast of roles (analyst, architect, developer, reviewer) and a ticket list, and run the roles in turn. Both are good at getting a better plan out of a model. This table is about what happens after the plan.
The rows are from reading their documentation and issue trackers in October 2026. Where codeloop is behind, the row says so.
The table
| Spec-first tools | Agile-persona tools | codeloop 0.4.1 | |
|---|---|---|---|
| Unit of work | A feature: one branch, one spec folder | A ticket in a list, with artifacts per initiative | A card with a story, in a lane. .codeloop/cards.json |
| The flow | Constitution, specify, clarify, plan, tasks, implement | Clarify, plan, build and verify, learn; ceremony sized after a look at the code | A lane file per kind of work: build, deploy, market, plan, triage, scan, analyze, learn |
| What decides a step is done | The agent, following the prompt. A shell script checks that the plan file exists | The agent writes the ticket's status | A command exits 0. done.cmd on every stage; the engine runs it |
| Where a person decides | In the chat, when the agent asks | A ticket stops one state short of done for a human to close | Named gates in the lane, with a role. Public steps stop before they run |
| Can the model skip a gate | Yes; gates are prompt text | Yes; status is model-written | No. An agent process cannot approve, and the engine refuses it |
| Board | None; a branch and checkboxes | An ordered ticket file | Cards, inbox, a web board, an optional cloud copy |
| Beyond the code | An idea-assessment extension | Modules for other kinds of work | Lanes for release, launch post, weekly plan, triage, competitor scan, growth review |
| Deploy | Not covered | Not covered; a human opens the PR | Staging and live stages, CI templates, a GitHub environment as the prod gate |
| Evidence | Verdicts from a bug extension | A review verdict per finding | evidence/<nnn>/verify.md with the commit SHA, per use case. An approval is bound to the output and the check; change either and the card parks again |
| A test that cannot fail | Not detected | Not detected | verify --mutate exits 4; lane eval refuses a changed check with no failing fixture |
| Learning | A tech-stack block in the agent context | A memory log per artifact | Wiki pages with frequency; critical at 3, promoted to rules at 10; lanes change through proposal, eval, promote |
| Unattended | A converge command inside one session | A loop command | run --agent from cron, with triggers, WIP and a per-agent run cap |
| Numbers | None | None | stats: human turns, unattended span, cycle time, rework, first pass per gate, per lane version |
| Priority | None; the order of the feature folders | The order of the ticket list | RICE on the feature, P1 to P4 bands by quartile; the inbox, card list --band and the Initiatives page sort by it |
| One SDK under CLI and UI | Slash commands only; no API | Slash commands only; no API | @protoboxai/codeloop/sdk: the CLI and the board call the same client, and each operation is tested through the HTTP and the local transport against one project |
| Hosts | Forty or more, through per-host command folders | Twenty or more, as generic skills | Claude Code, Cursor, Codex, and MCP |
| Ceremony for a one-line fix | The full pipeline | Sized after a short investigation | The full lane; an L card is refused at the spec gate until split. Size-adaptive lanes are planned |
| Cost per card | Not reported | Not reported | Not reported. Planned |
| Importing from the other | import speckit; import bmad reads an older file layout |
||
| Install | A Python tool plus uv |
A Node installer, or generic skills | npm install -g @protoboxai/codeloop |
| Age and community | Large | Large | Small. 0.4.1, one team, MIT |
What the table says
Spec-first and agile-persona tools help an agent write a better plan and then hand the code back to you. codeloop runs the whole loop, checks each stage with a command instead of a prompt, stops only at the gates you chose, and counts what happened. The founder's week is the proof for every row in the last column.
Where codeloop is behind: a trivial fix goes through every stage of the build lane, the cost of a card is invisible, three hosts is fewer than forty, and the BMAD importer reads a file its current releases no longer write. Each is on the Planned page.
Bringing work across
codeloop import speckit
codeloop import bmadOne card per existing feature or story. The source folders are only read. Lanes keep their checks, so an imported card goes through the same gates as any other.