loom

Agent Loom

Treat your coding-agent sessions like cattle, not pets.

Agent Loom makes the repo remember.

Loom is repo-local context integrity for long-running agentic software work.

Coding agents do not act from "the whole project." They act from the context they can see or retrieve, the tools they can call, and the verification they can run. Long work degrades when project truth is trapped in chat, lossy summaries, stale plans, scratch files, or one precious session.

Loom turns the stable mechanics of agentic coding into a Markdown-native truth graph:

• agents need shaped, retrievable context
• work needs bounded units and explicit contracts
• durable state should outlive the active session
• verification has to be externalized as evidence
• fresh workers should be safe to start, stop, replace, and hand off
• human and agent judgment should be preserved where future workers can inspect it

The result is simple:

• every durable claim has one owning record
• every fresh worker receives a bounded packet, not vibes
• every task closes with evidence, critique, and promotion when needed

The worker is disposable. The graph compounds.

Install Loom · Read the protocol · Architecture notes

Loom At A Glance

Read this as a noun/verb map, not an execution trace. The boxes are Loom nouns; the arrows are the verbs that keep agentic work decomposed into the right context owner.

flowchart TB
  classDef owner fill:#172033,stroke:#7d8590,color:#f0f6fc
  classDef ledger fill:#31243a,stroke:#b084eb,color:#f0f6fc
  classDef support fill:#2a2417,stroke:#8b949e,color:#f0f6fc,stroke-dasharray: 5 5
  classDef route fill:#163225,stroke:#7d8590,color:#f0f6fc

  B["loom-bootstrap<br/>load authority model"]:::route
  ROUTE["route by<br/>truth type"]:::route

  subgraph FRAME["Frame and intend"]
    direction LR
    C["constitution<br/>identity, principles,<br/>constraints, decisions"]:::owner
    I["initiative<br/>outcomes,<br/>success metrics"]:::owner
    SPEC["spec<br/>intended behavior,<br/>acceptance contract"]:::owner
    P["plan<br/>sequence,<br/>rollout, dependencies"]:::owner
  end

  subgraph KNOW["Learn and explain"]
    direction LR
    R["research<br/>investigations,<br/>tradeoffs, conclusions"]:::owner
    W["wiki<br/>accepted reusable<br/>explanation"]:::owner
  end

  T["ticket<br/>ONLY live ledger<br/>scope, blockers,<br/>acceptance, closure"]:::ledger

  subgraph JUDGE["Observe and judge"]
    direction LR
    E["evidence<br/>observed output,<br/>logs, tests, screenshots"]:::owner
    Q["critique<br/>findings, verdicts,<br/>residual risk"]:::owner
  end

  subgraph SUPPORT["Support and transport"]
    direction LR
    PKT["packet<br/>bounded child-worker contract"]:::support
    MEM["memory<br/>retrieval cues,<br/>preferences, reminders"]:::support
    EXT["external / support data<br/>PRs, issues, commands,<br/>logs, handoffs"]:::support
  end

  B -->|ask what changed| ROUTE
  ROUTE -->|principle / policy| C
  ROUTE -->|strategic outcome| I
  ROUTE -->|evidence gap| R
  ROUTE -->|behavior contract| SPEC
  ROUTE -->|sequence| P
  ROUTE -->|live work| T
  ROUTE -->|observation| E
  ROUTE -->|review| Q
  ROUTE -->|accepted explanation| W

  C -->|constrains| I
  C -.->|bounds behavior + work| SPEC
  C -.->|sets hard constraints| T
  I -->|frames outcomes| P
  I -.->|frames objective criteria| T
  R -->|informs tradeoffs| SPEC
  R -->|informs sequence| P
  R -.->|reduces uncertainty| T
  SPEC -->|contracts behavior| T
  P -->|orders bounded work| T

  T -->|delegates one slice| PKT
  PKT -->|returns for reconciliation| T
  T -->|requests observations| E
  E -->|supports / challenges claims| T
  E -->|grounds review| Q
  T -->|requests critique| Q
  Q -->|findings need disposition| T

  T -->|promotes settled learning| W
  W -.->|orients future work| T
  T -.->|promotes investigation result| R
  T -.->|promotes behavior clarity| SPEC
  T -.->|promotes sequence change| P

  MEM -.->|recalls graph pointers only| ROUTE
  EXT -.->|untrusted input| R
  EXT -.->|observed artifact| E

---

The problem

Most agent workflows eventually create a junk drawer.

PLAN.md becomes the spec, todo list, research log, failed-attempt record, review trail, status update, and handoff summary. Scratch files litter the repo. The active chat becomes a pet: overloaded with volatile decisions, painful to abandon, and weirdly valuable because the repo does not know what happened.

When the work stops, resumes, compacts, switches models, switches harnesses, or hands off to another worker, the next agent has to infer what is still true.

The model did not just forget.

The project never knew.

Loom fixes that by giving each kind of truth a home.

The idea

The active session is the wrong place for canonical project memory.