Skip to main content

Agent Catalog

ASDD defines 10 core specialized agents plus supporting agents for security, observability, and DevOps. Each agent has a precisely defined role, a minimum confidence threshold, and a failure protocol.


Core principle: separation of concerns

Each agent operates within a single, well-defined scope. No agent performs another agent's role. This separation is what makes the pipeline predictable, auditable, and recoverable when an agent fails.


The 10 core agents at a glance

AgentPhaseCore OutputMin ConfidenceFailure Action
Discovery Agent−1 / 0Behavioral slices from raw intent0.85Draft mode; route to PO/TL
Spec Agent1EARS-format requirements0.85Flag ambiguous sections; route to TL
Validation Agent1Spec quality score + validation report0.90Block pipeline; TL sign-off
Domain Agent2Schema-compliant domain model0.85Draft mode; TL review
Design Agent3Architecture + ADRs + Reasoning Trace0.80Draft architecture; TL review before proceeding
Task Planning Agent4Execution wave plan (tasks.md)0.80Draft plan; TL review
Implementation Agent4Code via context-fresh sub-agents0.75Feature branch; human code review
QA Agent1 + 5Test suites mapped to spec behaviors0.85Flag uncovered requirements
Refactor Agent4Refactored code within spec boundaries0.80Flag; TL review
Knowledge AgentAllState manifest, learning proposals0.80Propose update; human approval

Supporting agents

AgentRoleMin Confidence
Security AgentPre-deployment compliance scan0.95
Observability AgentTelemetry instrumentation and validation0.85
DevOps AgentCI/CD pipeline automation0.85

Agent interaction pattern

Agents interact through the State Manifest, not directly with each other:

This indirect interaction pattern:

  • Prevents agents from directly overriding each other's outputs
  • Creates an immutable audit trail of every state transition
  • Allows the Knowledge Agent to detect conflicts before they cause cascading failures

Context injection: what agents receive

The Knowledge Agent is responsible for Dynamic Context Injection — ensuring each agent receives only the context relevant to its current task.

At each phase gate, the Knowledge Agent constructs a minimal context payload containing:

  1. The active slice's phase_data (links to intent, requirements, design artifacts)
  2. The applicable Ubiquitous Language terms for the current domain area
  3. The Steering Rules relevant to the current phase
  4. Prior Reasoning Traces from upstream agents in the same slice
  5. Any open dissent notices or blockers for the slice

Everything else is withheld. A clean context window is as important as clean code.


Reasoning traces

Every agent emits a reasoning_trace.md alongside its artifact. The trace documents:

  • Key assumptions made
  • Requirements prioritized and why
  • Alternatives considered and rejected
  • Uncertainty factors (required when confidence < 0.95)

Phase gate sign-off requires the TL to acknowledge the Reasoning Trace — not just the artifact.


Each agent has a dedicated page with full specification: