Your agent ships a feature and it looks right. A week later someone asks which requirement it covered, what got ruled out, and what proved it worked. The answers are in a chat you already closed.
I kept ending up there, so I built agentsmyth. The name is two halves. The myth is the one I set out to break: that a model, an agent, and an agent working inside a harness are the same thing and behave the same way. The smyth is the forge, because a harness is only worth having if it takes the heat and keeps its shape.
I did not expect most of the myths it broke along the way to be mine.
I thought writing the code was the job
Most of the code in my repos is no longer typed by me. An agent writes it faster than I ever could, and somewhere along the way I stopped trying to compete with that.
What surprised me is which parts of the work got heavier. Planning, reviewing, testing. The parts I used to rush, or quietly skip when a deadline got close, are now most of what I do.
The lazy parts turned out to be the job.
I find that more interesting than threatening. When code arrives faster than anyone can read it, the process around it has to carry the weight, and that process is still mine to design. So that is what I built: a harness. A model only answers, and an agent acts on the answer. A harness is what decides when the agent is allowed to call something done.
harness (agentsmyth)
┌──────────────────────────────────────┐
│ think → plan → build → review → │
│ test → ship → reflect │
│ ┌──────────────────────────────────┐ │
│ │ agent (Claude Code, Codex, ...) │ │
│ │ ┌──────────────────────────────┐ │ │
│ │ │ model │ │ │
│ │ └──────────────────────────────┘ │ │
│ └──────────────────────────────────┘ │
│ git hook: no approval, no commit │
└──────────────────────────────────────┘
I thought the files were the harness
Earlier this year I wrote seven skill files to keep an agent on track through a single refactor, one per phase: think, plan, build, review, test, ship, reflect. I documented that run in a nine-part series called The Contract, and it worked.
With me approving every handoff.
The files were not the harness. I was. Every brief, every plan and every verdict passed through me before the next phase could read it, and nobody installing this in their own repo gets me standing at their gates.
So the rules had to move somewhere that does not need a person or a prompt to remember them. They moved into git. agentsmyth sits on top of the agent you already run, whether that is Claude Code, Codex, Copilot, Cursor or Windsurf, and a hook checks every commit against the trail each phase leaves behind. A plan that nobody approved does not get past it.
$ git commit -m "plan: session handling"
build: workflow/artifacts/plans/add-session-handling-v1.md → ready-for-next-phase ✓
check-lifecycle --phase build --slug add-session-handling: failed with 1 issue(s)
$ echo $?
1
The plan marked itself ready. Git checked who had approved it, found nobody, and stopped the commit.
The agent cannot talk its way past an exit code.
Small changes stay small, and a typo fix never needs a plan. Everything else leaves a written record from each phase that the next session reads instead of a chat it never saw. I built every release of agentsmyth this way. Its own repo carries 38 briefs, 38 reviews and 34 retrospectives across five releases since July.
I thought more agents meant more coverage
Version 1.1.0 added councils. On complex work, several agents research and review in parallel, and one of them is assigned to challenge the rest.
Before shipping it, I measured it against a single agent doing the same research. The single agent found more, for a fraction of the calls. Not the result I wanted.
I shipped the council anyway, because of the challenger. In that run it refuted a wrong claim from one of the researchers. The single agent had nothing checking its own claims.
Agreement between agents is not evidence.
So no council can approve a phase on its own. It raises findings, and the agent running the phase has to accept, merge or reject each one with a reason.
I thought a strict tool would be strict with itself
Across 154 issues the reviewers raised so far, every one was accepted or merged. Not a single rejection.
The tool refuses my commits without hesitation. It has never once refused its own council.
That is either a very good council or a very polite one, and I do not have a way to tell yet.
The paywall I planned and never built
Early on I planned a paid tier for some of the skills. Then I looked at what already existed. GitHub's Spec Kit, BMAD-METHOD, claude-task-master and agentpreflight all give agents structure, and all of them are free. A skill is also just Markdown an agent has to read in plain text, so locking it buys friction and nothing else.
Every skill ships free, and that is a decision, not a placeholder.
If your real problem is vague specs, Spec Kit goes deeper there, and I say so in my own README. agentsmyth is for the other problem: the agent skipped the plan and shipped anyway.
The quickest way to test any of this is to run it
The current release is 1.1.1. You need Node 20 or later and a git repo.
npx @jeelvankhede/agentsmyth@latest init
# then, inside your agent, in the same repo
/agentsmyth # Codex: /prompts:agentsmyth
init sets up the folders and the git hook. The slash command finishes setup: your agent reads the repo and asks only what it cannot work out. Use the command rather than asking in plain English, because an agent can skip a casual instruction but not a typed command.
When a new version lands, npx @jeelvankhede/agentsmyth@latest upgrade brings your repo current and backs up anything you edited first.
If you run more than one agent on the same change, look at how often one of them actually says no. I would like to know what your number is.
What you can run today:
agentsmyth adds a seven-phase lifecycle to the repo you already have, works with the agent you already use, and enforces the order at
git commit. Install it withnpx @jeelvankhede/agentsmyth@latest init. What it leaves behind is a readable trail for every change that mattered. Free and MIT licensed.
Top comments (0)