AI agents

ARTA is not one prompt in a loop. It is a pipeline of specialized agents — plain Python classes under src/agents/ — each owning one responsibility, with deterministic validators between the LLM and anything that ships.

The core pipeline agents

Agent	Module	Responsibility
Orchestrator	orchestrator.py	Coordinates the pipeline end to end
Strategy architect	strategy_architect.py	Requirement analysis and risk scoring (probability × impact)
Requirement decomposer	requirement_decomposer.py	Splits requirements into testable units
ATDD designer	atdd_designer.py	Gherkin acceptance criteria per requirement
Automation engineer	automation_engineer.py	Generates test scripts per runtime, grounded in discovery context
Grounding validator	grounding_validator.py	Rejects hallucinated selectors, roles, and endpoints at generation time
Execution agent	execution_agent.py	Drives test execution
Defect intelligence	defect_intel.py	Classifies failures (sut_regression / test_gen_bug / grounding_blocked) and files defects
Quality gate	quality_gate_agent.py	Aggregates results into a pass/fail gate
Self-healing	self_healing.py	Consumes the regeneration queue; retries failed generation with corrective hints

Discovery and understanding agents

Agent	Module	Responsibility
API discovery	api_discovery.py	Captures real endpoints from the running SUT
Architecture discovery	architecture_discovery.py	Builds a picture of the SUT’s structure
Discovery executor	discovery_executor.py	Runs the read-only discovery probe
Protocol discovery	protocol_discovery.py	Classifies endpoint protocols beyond plain REST
Endpoint grounding	endpoint_grounding.py	Maintains the store of verified-real endpoints
GitHub context	github_context.py	Pulls route/DTO context from the SUT’s source repository
DTO extractor	dto_extractor.py	Derives request-body shapes from SUT source

Supporting machinery

Around the agents sit deliberately boring, deterministic pieces: retry ladders and retry policies, a circuit breaker, LLM client abstractions (the Ollama client and cloud-provider clients), traceability agents that write the graph, and evidence collectors that keep reports auditable. The LLM proposes; deterministic code disposes.

The design position: an agent may be creative only inside boundaries a validator can check. That is what makes the difference between “AI wrote my tests” and “AI wrote my tests, and here is the proof they are real.”