Agents

Agent-native source control without operational chaos

A machine-speed team needs source-control primitives that let many actors rebuild software continuously while humans retain governance over the loop.

Agents are first-class actors

LLMs, copilots, bots, CI workers, and deployment controllers act against the same graph with predictable conflicts, leases, and policy checks.

Workspaces stay cheap

Centralization does not remove branching; it enables high-velocity branch, fork, and overlay operations with shared policy context.

Humans keep authority

Folder-level, path-aware, and automation-level constraints keep sensitive areas protected while agents do the repetitive coordination work.

What Git never had to model

Workspace leases for parallel agents that need temporary authority over a slice of the graph.

Policy-aware autonomy levels so observe, recommend, guardrailed action, and autonomous remediation stay distinct.

Runtime signal events that can trigger a source change, not just page a human.

Replayable operation history that lets humans inspect why an automated loop advanced or stopped.

Agent operating loop

  1. 01

    Agent reads source, runtime, cache, and policy state through the same control plane used by humans.

  2. 02

    It creates an isolated workspace or overlay instead of mutating shared truth directly.

  3. 03

    Policy checks resolve path, actor, tenant, and automation-level constraints before risky operations.

  4. 04

    Commit, operation-head, and ref updates are validated with CAS semantics to avoid silent overwrite.

  5. 05

    Automation output is reviewable and traceable as first-class graph events.

  6. 06

    Runtime feedback decides whether the loop advances, retries, reverts, or escalates to a human.

Integration focus

Beyond human-centric PR queues

Agents can drive patch, publish, stack, and canary flows without treating a manual pull-request queue as the only path forward.

CLI-native agent orchestration

By being JJ-native, tooling can expose operation invariants directly instead of hiding them behind Git adapters and shell scripts.

Auditability

All high-risk changes should be explainable from actor identity, remote state transitions, runtime signals, and operation history.

End-state goal: humans govern the loop, while agents execute the high-throughput operational workload continuously, reproducibly, and with evidence.