Advisor
Clarifies intent, prepares acceptance criteria and selects how the spec should run.
↳ Spec + decisions
HUMAN INTENT. AGENT EXECUTION. RECORDED EVIDENCE.
Intent in.
Evidence out.
Build with your coding agent.
Keep the implementation and intent.
Choose how much to delegate.
AIDA / OPERATING MODEL
One system. Multiple agents. Explicit accountability.LINUX · GIT-NATIVE · MULTI-VENDOR
01 / THE SHIFT
Writing code got faster.
Engineering still needs a system.
People and coding agents build together. AIDA helps preserve what the system should do, why decisions were made, and how that intent connects to the implementation.
AIDA is a governed software factory with selectable autonomy. It gives agents shared requirements, defined roles, independent review and enforceable limits—so a team can delegate the work and retain control of the outcome.
02 / INSIDE THE FACTORY
A repeatable path. A durable record.A team of agents.
A system of responsibility.
AIDA coordinates the lifecycle around the coding agents you choose. Each role has a job. Each handoff carries context.
Implementer
Works against the spec, changes code and validates the result in an isolated worktree.
↳ Change + tests
Reviewer
Independently checks the change against its acceptance criteria and records a verdict.
↳ Verdict at a commit
Integrator
Lands approved work through the applicable CI, review and authority gates.
↳ Gated merge
✳
Shared intent survives the handoff. Specs, relationships and history live in a Git-backed store. Operational records preserve the evidence around reviews and merges.
03 / SELECTABLE AUTONOMY
ONE PROJECT. DIFFERENT LEVELS OF DELEGATION.Choose where
people stay in the loop.
Run a dark factory for bounded work. Keep people close to consequential decisions. Choose the execution mode per spec.
DARK FACTORY MODE
Delegate the lifecycle.
Agents implement, test, independently review and merge approved, authorized work. People set the intent, acceptance criteria and limits. Exceptions escalate.
A GOOD FIT
A bounded fix with clear acceptance criteria and no unresolved design decision.
WHO DOES WHAT
- Define intent & limitsHuman
- Implement & testAgent
- Independent reviewAgent
- Merge through gatesAgent
Routine human coding and code review are removed from the execution path. Review, CI and authority gates remain.
Some work needs a decision first. decide blocks execution until the pending decision is resolved. It sits outside the autonomy ladder.
04 / THE CONTROL LAYER
BEYOND A PROMPT AND A PULL REQUESTBuilt for the work
between the agents.
Intent that persists.
Versioned requirements, decisions and relationships give the next agent a place to start. Trace comments connect code to the spec that justified it.
Requirements · Decisions · TraceabilityParallel, with boundaries.
Queues, worktrees, leases and role-specific sessions coordinate multiple workers. Mail and handoffs carry context across session boundaries.
Queues · Worktrees · SessionsReview with a record.
Independent reviewers assess acceptance criteria. Verdicts attach to exact commits, and merge gates check the evidence and applicable holds.
Review · CI · Merge gatesDelegation with limits.
Seat grants and time-boxed mandates define authorized actions. Work outside delegated authority waits for the appropriate decision.
Grants · Mandates · Escalation05 / THE MANAGEMENT CASE
Delegate more.
Keep the evidence.
AIDA makes the operating model explicit: what is being built, who is responsible, and what allows it to move forward.
Read the executive brief · Engineering & DORA auditMake human attention deliberate.
Keep design decisions interactive. Let well-specified work proceed through the autonomous lifecycle.
Inspect the basis for a change.
Follow a requirement into code, its review verdict and the record of how it progressed.
Keep your choice of coding agent.
Coordinate Claude Code, Codex CLI and Antigravity around the same spec graph and operating rules. Harness capabilities and containment support differ.
06 / START WITH ONE CHANGE
Give the next agent
more than a prompt.
Choose a bounded change. Write down what success means. Run it through AIDA and inspect the resulting evidence.
Explore capabilities · Connect project tools · Configure your operating model · Follow the workflow · Inspect trust and evidence · Explore Metrics · Engineering & DORA audit · Browse the library · Complete product assessment
Available today: alpha, Linux only. AIDA operates under a shared Unix user; multi-human identity separation is not a security boundary. Production deployment needs its own policy and automation.