AIgency · AgentDock Pro
Describe the outcome. AIgency runs the work.
AIgency turns one objective into a plan of tasks, gives each task to the agent and model that suit it, runs them in order — and in parallel where it can — and checks the result before it calls the job done.
AIgency · clinic-booking
Add an SMS opt-out to reminders
- Add an opt-out flag to patientsCodex · GPT-6 Astracompleted
- Skip opted-out patients in the reminder jobClaude Code · Opus 5.5completed
- Opt-out toggle in the web appAntigravity · Gemini 3.1 Procompleted
- Tests for opt-outClaude Code · Fable 5.1completed
- Repair iteration 1: Tests for opt-outClaude Code · Fable 5.1completed
- Review the whole changeCodex · GPT-6 Solcompleted
How a run works
A plan you can read, then a run you can watch.
01
Describe the outcome
One objective in plain words, a project folder, which agents may take part and how thorough to be.
02
Review the plan
A planner drafts tasks, dependencies and checks. It's a proposal: edit, reorder or reassign before anything runs.
03
It runs the work
Each task is an ordinary AgentDock session. Tasks start when what they depend on is done; independent ones run in parallel.
04
It checks, repairs, reports
Checks run against the result. A failure gets a bounded repair loop; the run ends with a report of what was done and what wasn't.
Routing
Each task goes to the model that suits it.
AIgency first rules out models that can't do a task, then ranks the rest on imported benchmark evidence, and runs the winner on the agent that launches it. Models marked manual-only are never chosen automatically, and retired ones are never chosen at all.
Lean the balance your way with a routing preset:
- Balanced
- Claude Heavy
- Codex Heavy
- Visual / UI Heavy
- Efficiency
- Custom
Model-based routing ships in the next release; the current release routes by agent.
Quality gates
Done means checked.
Attach acceptance criteria to a task or the whole run. When one fails, AIgency starts a bounded repair — fix, then check again — instead of reporting success it hasn't earned.
- Tests pass
- Build succeeds
- Typecheck succeeds
- Lint succeeds
- Command exits successfully
- Output contains a value
- Output must not contain a value
- Required files exist
- Artifact produced
- Integration completed cleanly
- Reviewer approves
- I will approve this manually
You stay in charge
Autonomy, at the level you choose.
Pause, resume, stop
Pause a run between tasks, resume it later, or stop it. Waiting on capacity, it backs off and tells you why.
Approval levels
Ask before actions, auto-approve routine work, or full autonomy — and tasks can require your sign-off before they finish.
Separate worktrees
Tasks can run in their own git worktree so parallel work never edits the same files at once, then integrate.
Start from a template.
Seven built in, and any run can be saved as your own.
- Build a feature
- Fix a bug
- Review a repository
- Refactor code
- Investigate a technical problem
- Plan and implement an MVP
- Test and validate an implementation
End with a report.
Every run closes with a completion report you can export.
- What was done
- Files changed
- Checks
- Commands run
- Repaired
- Still unresolved
- Worth knowing
Or just ask for it in a conversation.
Type @aigency and an objective, or let the agent you're talking to suggest it. You get the plan to review in the conversation, and nothing starts until you do.
AIgency is part of AgentDock Pro.
Pro is A$29.99 a month or A$299.99 a year, held on your Sarja Account. Orbit, the desktop app for Antigravity, is included.