Run an organization as a folder of plain files — operable by a human with a text editor, accelerated by any AI coding agent, owned by neither.
Traditional software freezes an organization's workflows into applications. Now that software can be generated and modified by agents, the durable asset is no longer the app — it is the organization's own state: purpose, authority, processes, work, evidence, decisions, and lessons. This seed keeps all of that as Markdown + YAML in a git repo, and treats every agent, model, script, and scheduler as replaceable machinery.
Your business, as a folder any AI agent can run — and none can run away with. Software used to freeze your workflows into apps; now agents can generate software, so the durable asset is your organization itself: purpose, authority, processes, evidence, decisions. The seed keeps all of it in plain files under git. Agents are hired like staff — chartered roles with narrow grants, hard boundaries, probation — and fired by deleting a file. You stay the judge: nothing spends, sends, signs, or files without you. Start with one painful recurring process, not a migration; your existing tools keep running. It's proven, not theoretical: a real eight-figure (INR) business runs on it today — books reconciled to the paisa, every marketplace refund provably credit-noted from the right account, support drafted every morning before the owner wakes up. Fork the seed. Keep the judgment. Delegate the rest.
The thesis: the seed is the genotype; everything else is phenotype. Harnesses, models, and SaaS systems are expression machinery — they improve every quarter, and because the organization's identity lives in files and not in any of them, every improvement anywhere in the stack is captured as a free upgrade. The org rides every curve and is owned by none. Roadmap: docs/ROADMAP.md.
Origin: extracted from a living instance that runs a real e-commerce business — its purchase orders, support desk, books reconciliation, and governance all flow through this exact structure daily. The pattern is published; the business stays private.
v0.2 folds back what sustained live operation taught: chartered agent Roles
(roles/_charter-template.md — narrow grants, hard boundaries, probation), the
concurrency convention (session leases + fencing + deadman checks; work/_active/),
two-tier governance (full Proposals only where power grows; decisions/fast-track.md
for everything else), and the write discipline that makes crashed or colliding agents
recoverable by design (docs/write-discipline.md). All of
it was forced by real failures — concurrent agents in one ledger, mid-write crashes,
silent schedule deaths — not designed in advance.
An organization of any size is a quantity of eight kinds of durable thing — scale adds files, never new kinds:
| # | Category | Lives in | What it is |
|---|---|---|---|
| 1 | Identity | ORG.md |
purpose, current goal, the Now |
| 2 | Roles | roles/ |
positions with responsibilities + authority; humans or agents occupy them |
| 3 | Processes | processes/ |
how kinds of work are done — steps, mechanical Checks, one Judgment question |
| 4 | Tasks | work/ |
one bounded performance of a process |
| 5 | Records | records/ |
the nouns the business touches; each Kind defined by a _kind.md (this is your ontology) |
| 6 | Decisions | decisions/ + inline |
recorded exercises of authority |
| 7 | Lessons | lessons/ |
preserved experience; every process reads its lessons first |
| 8 | Proposals | proposals/ |
the mutation queue — reviewable, reversible changes |
…governed by AUTHORITY.md (the rulebook), with Mounts (AGENTS.md,
CLAUDE.md, .claude/, connectors) as disposable harness bindings outside the
stencil entirely.
Every piece of work travels the same path:
- A human expresses intent (ephemeral — never a filed artifact).
- The intent becomes a bounded Task under a Process.
- A human or agent performs it; every claim in the output cites Records.
- Mechanical Checks verify coherence (no LLM required); a human rules the Judgment.
- Lessons are recorded; frictions that would change a governed file become Proposals — applied only after a human ruling, always with a rollback.
Intent with no matching process is not an error — it is the growth signal: the operator drafts a proposal for the new process, and the organization grows one approved mutation at a time (directed evolution — see docs/adr/0002).
- Bindings, not homes. Skills, agent files, and connectors point at durable state; they never own it. Delete every mount and the organization still runs. (ADR 0001)
- A small conserved core. Identity, authority, roles, and processes change only via approved Proposal. Everything else is free.
- Access is not permission. A connected tool grants nothing; only
AUTHORITY.mddoes. External writes and spending always stop at a human boundary. - Evidence is a citation discipline. A claim without a citation is unverified by definition; checks are runnable by hand.
- Systems are federated, not migrated. Gmail, your ERP, your helpdesk keep owning their data; the seed records who owns which truth (records/systems/) and governs how it may be touched.
- Lessons weigh what their evidence weighs. Recurring evidence upgrades a lesson; a closed one-off decays. Scar tissue is audited, not accumulated.
- Use this template (GitHub → "Use this template", or clone).
- Write your
ORG.md— fill every{placeholder}: purpose, a Founder and an Operator role, an honest Now section. - Keep
AUTHORITY.md's reserved powers (they travel well verbatim); write your Operator grants. - Enumerate your external systems as
records/systems/entries — what each is the source of truth for, what access exists. This moves no data; adoption is acknowledgment, not migration. - Write one process for one real recurring pain, using processes/example-weekly-review.md as the shape: steps a human could follow, Checks, one Judgment question, "read the lessons first" as step 1.
- Run the loop once — intent → task → evidence-cited output → checks → your judgment → lesson → proposal. Then keep running it.
- Point your agent at it. Any coding agent that reads
AGENTS.md(orCLAUDE.md) lands inORG.mdand knows the org, its authority, and the next action. Switch harnesses any time — the folder is the organization.
Don't migrate. Federate. Your systems keep running; the seed starts as the governance layer: enumerate the systems, run one process against them, and let records migrate inward only when a process earns it. A business is "on the seed" the day one real decision flows through the loop — not the day its data moves.
When you find doctrine trapped in old tooling (playbooks, prompts, skills), migrate it curated: read everything, take what earns its place, leave the wrapper, record the gaps you find. The conserved home outlives every tool that visits it.
ORG.md ← your canonical entry (template)
AUTHORITY.md ← the rulebook (template; reserved powers ready)
CONTEXT.md ← the glossary of seed terms (keep it)
AGENTS.md / CLAUDE.md ← thin mounts for any coding agent
processes/example-weekly-review.md ← a worked example process
records/systems/_kind.md ← the Kind that makes adoption = acknowledgment
proposals/0000-proposal-template.md ← the mutation form (reason, evidence,
benefit, validation, rollback, ruling)
docs/adr/ ← why the pattern is shaped this way
Directories like work/, lessons/, decisions/ appear when your first task,
lesson, or decision does — the seed never ships empty scaffolding.
Apache-2.0. Fork it, run a company on it, build products above it.