What is governed autonomous development, and how does Arcadia approach it?
In brief
Governed autonomous development lets AI coding agents advance bounded work within authority defined by a human. Durable intent determines priority, consequential decisions stop for approval, and observable evidence establishes completion. The human owns direction; agents carry approved work forward across sessions.
- Define authority before agents act.
- Preserve decisions across sessions.
- Require evidence before claiming completion.
The problem: autonomy without durable boundaries
Governed autonomous development lets agents advance bounded work within human-defined authority. Reviewing every mechanical step consumes attention; unlimited execution can carry a mistake into spending, publication or production.
The challenge is keeping three things clear across sessions: what the work serves, which decisions remain yours and what proves completion. Instructions alone cannot enforce access controls.
What you may have tried
- Approving every step: keeps you involved, but interrupts routine progress and does not scale well.
- A broad “keep going” prompt: grants momentum without preserving specific approval boundaries.
- CI and written rules: useful foundations, but a passing build cannot authorize deployment, and rules need enforced permissions.
Separate reversible execution from consequential decisions. Define acceptance criteria before coding and keep approval ownership explicit.
How Arcadia connects autonomy to authority and proof
Arcadia maintains plans and actions with one work pointer, records decisions and approval gates, and checks completion evidence against criteria written before work starts. A Project’s Mission explains why it exists; its Outcome defines the desired change. Milestones mark progress, and a Plan organizes Actions small enough to finish and verify. Decisions and evidence remain connected to that work.
For a signup improvement, the Outcome might be dependable registration. An Action fixes invalid-email handling and records checks against the candidate revision. The agent can prepare and verify the change while production release still waits for authorized approval.
This preserves useful autonomy without treating a successful check as permission. Human priorities, access controls and risk acceptance remain essential; Arcadia’s development approach is not a compliance certification. Start by writing what one Action may do independently and what must stop for you.
Sources and further reading
- Anthropic: building effective agents — choosing bounded workflows and evaluating agent behavior.
Current capabilities and supporting evidence
- Plans, actions and a work pointerWorking
Governed documents with one active plan and one current action.
Supporting evidence - Decisions and approval gatesWorking
Consequential calls are recorded and wait for a human answer.
Supporting evidence - Completion evidenceWorking
Acceptance criteria are checked before an action can be marked done.
Supporting evidence
Capability snapshot: . See what changed →