Product

Version control rebuilt for teams where agents commit alongside humans.

The record your team leaves behind should carry the reasoning, not just the code. Every thread isolated, every action attributed, every state permanent. Loom is one coherent system where human and agent work share a single legible record.

01

Context annotations

The reasoning travels with the code.

Constraints, rationale, and design intent belong next to the code they govern — not in a wiki, not in a Slack thread, not in the session transcript of an agent that has already expired.

Annotations are structural. Each one is scoped to a file, a symbol, or a line range, and carried forward on every snapshot that touches its anchor. When the next agent opens the same function, it sees the reasoning before it starts editing.

Foundations for annotations ship today in the object model; the editing surfaces are on the roadmap. Clearly labeled throughout.

02

Threads

Many agents. One repo. Zero collisions.

A thread is a named pointer into the DAG. Opening one creates an isolated worktree, sharing the main repository's object store. Many threads, zero extra disk. No coordination, no merge dance before the work is ready.

Threads track lineage structurally. A rebase, a collapse, or a merge produces a new state without overwriting the original — so another agent that forked earlier on the same thread keeps its view intact.

$ loom start feature/auth --agent claude-opus-4.7
thread started · worktree: .loom/agents/claude-opus-4.7

$ loom start feature/auth --agent claude-sonnet-4.6
thread started · worktree: .loom/agents/claude-sonnet-4.6

# two agents, one repo, no conflict.
03

Attribution

You always know who acted. Human or AI.

Every state carries first-class attribution: provider, model, confidence score, verification signature. Never parsed from a commit message. Never inferred from a Git email. Recorded at the moment the action landed.

When an auditor — a CTO, a staff engineer, or a compliance tool — asks who touched this code, Loom returns a structural answer, not a best-guess string.

04

History model

Nothing disappears. Every object permanent.

States are content-addressed with BLAKE3. IDs are stable across clones, forks, and time. Rebase and collapse create new objects; the originals survive, reachable from whatever reference held them.

The kind of history you can point to a year from now and trust completely — not because you remember what you did, but because the record is structural.

05

Governance

Policy at the namespace. Legible at every level.

Namespace grants cascade. Repository grants scope precisely. When access fails, you see which layer said no — token scope, namespace policy, or repo grant — not an opaque 403.

Effective permission is always explicit. You never guess what a token can do; you read it.

One coherent surface

Quiet enough for leadership. Precise enough for staff engineers.

The same page that gives a CTO the trust summary gives a staff engineer the attribution trail, the thread lineage, and the signing material. One record. No reconciliation across source host, CI dashboard, and chat log.