Auto Office

Intent: make agentic software development behave like a disciplined engineering team, not a single coding agent with a long prompt.

Auto Office is an opinionated software-development lifecycle for AI agents. One orchestrator owns strategy; specialist agents plan, implement, review, verify, and close out; the office runtime owns the mechanical state transitions that make those roles reliable.

The goal is simple: take a human intent and turn it into a reviewable, resumable, evidence-backed engineering run with explicit authority boundaries.

• Distribution: auto-office
• Python package: office
• CLI: office
• State authority: SQLite runs.db (WAL)
• Core rule: agents decide; the runtime records, isolates, routes, verifies, and resumes

Auto Office is deliberately not a generic multi-agent chat framework. It has a lifecycle, hard invariants, role-specific routing, isolated worktrees, immutable submissions, independent review, integration gates, version-pinned runs, and explicit landing authority.

New XS–XL run gears: Funded behavior, subtask sizing, adaptive profiles, Goldfish and legacy compatibility.

Start here: Manifesto · Install · Opinionated SDLC · Routing · Benchmarks · State & resume · Development

Auto Self Improve (issue-only)

When an outcome looks like an Office defect, the runtime prints one non-blocking bug?: line (a refusal of the command next: names, a failed job or gate, a lost dispatch, an uncaught internal error). It is a hint: the runtime starts no investigator, makes no model or network call and files nothing.

The orchestrator decides whether the hint is an Office defect, dedupes against open and closed issues, sanitizes the report and files it with gh issue create --repo Hikari9/auto-office, directly or through a cheap read-only subagent. Only Auto-Office-owned defects are candidates; target-application bugs and routine gate refusals are not. office self-improve is a read-only pointer: it prints these steps and the run’s current bug?: signals. Issues only: nothing here fixes a bug or opens a PR, and filing never blocks landing. The workflow is in skills/auto-self-improve/SKILL.md.

The lifecycle at a glance

flowchart LR
    A["Human intent"] --> B["Tracking issue + office start"]
    B --> C["Discovery + plan"]
    C --> D["Plan diagramWaves · routes · checkpoints"]
    D --> E{"User approval"}
    E -->|approved| F["Parallel / stacked execution"]
    F --> G["Immutable submissionchecks + executor self-review"]
    G --> H["Lane composed: one convergence review+ visual review when applicable"]
    H -->|RECHECK| F
    H -->|INTAKE_GAP| C