Playbook › Build: agent development
Build: agent development
The build phase: agent development through the ADLC. The four tracks hand their outputs to one
disciplined pipeline. The differentiating construct is the mini-output: every agent produces a bounded, independently reviewable piece of the
final deliverable, and only the orchestrator assembles the whole. No single agent ever writes the final artifact.
Step by step
Enter the pipeline
- ADLC entry. One-page spec plus an architecture decision record per capability, referencing its PRD-evals from Track 2 and its pattern (P1 to P7).
- Split per the determination rule. Everything with a testable right answer goes to the domain rule library as deterministic code; only judgment steps become agent work. The split was drafted in Tracks 2 and 3; here it is enforced by review.
- Author to the SSRA contract. One skill per use case; agent definitions carrying the four-part delegation spec (objective, output format, tools and sources, boundaries); run contracts with a machine-checkable end state, a token budget, and effort scaling for the job class.
Mini-outputs and assembly
- Mini-outputs. Each agent's run contract names the mini-output it owes: a section, a reconciled table, a risk register entry, a cited analysis. Mini-outputs are structured content, small enough to review in minutes and to regenerate cheaply.
- Orchestrated assembly. The orchestrator collects mini-outputs, checks completeness against the dependency graph, and assembles the final deliverable through the deterministic rendering boundary: immutable template plus approved structured data. The orchestrator submits; it never authors.
Quality engineering
- Guardrails on every output. Every mini-output passes deterministic checks before it is eligible for assembly: schema validity, citation presence, no unapproved values, classification respected.
- Loop engineering. Each mini-output is scored against its evals; a failing output regenerates with the eval verdict fed back, bounded by the run contract's budget; persistent failure escalates to a person rather than burning budget.
- Automated unit tests as evals. Every agent is mapped to its capability and target metrics; the PRD-evals run as pre-deploy gates in CI, alongside the deterministic test suite. Nothing ships scoring below its predecessor, and nothing unregistered can reach client data.
- Integration wiring. The parametrized requirements from Track 3 are configured, never coded: connections, contracts, mappings, and schedules land as plane values and the pipelines run from templates.
Exit criteria per capability
| Check | Evidence |
|---|---|
| Evals pass at target | PRD-eval scores at or above threshold on golden exemplars |
| Tests green | Deterministic suite plus tool-layer evals in CI |
| Human review configured | HITL placement, approvers, and escalation timers set in the plane |
| Pattern and antipattern review | Build maps to one of the seven patterns; antipattern checklist clean |