Architecture
Five layers. One continuous loop.
An autonomous organization is structured as a stack of five layers, connected by a continuous loop. Goals flow down. Verified outcomes flow up. Knowledge feeds back into everything.
The continuous loop
DEFINE
Set goals with measurable criteria
DISPATCH
Match agents to milestones
EXECUTE
Agents produce outputs
EVALUATE
Binary verification of every output
LEARN
Capture knowledge, detect patterns
The stack
Each layer has a single responsibility.
Goals, not tasks
YAML-defined goals with milestones and success criteria. Each milestone has binary pass/fail criteria -mechanical checks and agent judges. The goal layer is the single source of truth for what the organization is trying to accomplish.
Key properties
- Goals decompose into milestones with explicit dependencies
- Each criterion specifies its verification method: test, file check, or agent evaluation
- Progress is measured by criterion pass rate, not time spent
Agent orchestration
A scheduler evaluates which milestones are actionable, selects the right agent for each, and dispatches with full context -specs, prior attempts, runtime environment. Agents don't decide what to work on. The system does.
Key properties
- Dispatch includes spec, acceptance criteria, runtime context, and prior failure analysis
- Agents are matched to milestones based on role and capability
- Budget limits prevent infinite retry loops -escalation kicks in after exhaustion
Specialized agents
Workers receive a directive and produce output. They write code, generate docs, run tests, review architecture. Each agent has a defined persona, toolset, and authority boundary. They don't coordinate with each other -they report back to the dispatch layer.
Key properties
- Executors implement. Reviewers verify. Architects evaluate system-level decisions.
- Worker interviews capture post-mortem observations after every dispatch
- Agent specialization prevents the 'one model does everything' failure mode
Binary verification
Every output is evaluated against its criteria. Mechanical criteria run automated checks -file existence, test pass rates, type safety. Subjective criteria dispatch evaluator agents with explicit rubrics. There is no 'looks good to me.'
Key properties
- False passes are tracked and fed back as anti-patterns
- Evaluation is independent of execution -different agents, different context
- Criteria that can't be verified mechanically are flagged at goal definition time
Learning and memory
Completed goals produce structured knowledge. Worker interviews are synthesized. Patterns are detected across dispatches. Evaluation data reveals what approaches succeed. This knowledge compounds and informs future work.
Key properties
- Worker interviews capture friction, workarounds, and improvement suggestions
- Pattern detection identifies recurring strategies and failure modes
- Knowledge is queryable and automatically injected into future dispatch context
Agent roles
Specialized agents, clear boundaries.
Each agent has a defined role, capability set, and authority boundary.
Executor
Writes code, generates content, produces artifacts. Receives a spec and returns an output.
Reviewer
Evaluates outputs against criteria. Independent of the executor. Flags issues, not fixes.
Architect
Evaluates system-level decisions. Reviews specs before dispatch. Catches structural problems early.
Orchestrator
Sequences work, manages dependencies, handles scheduling. Doesn't execute -coordinates.