Note: agents

Context

The agents module implements an AI-powered release preparation system that automates code quality assessment and release readiness validation.

Agent architecture

The system uses a progressive tier escalation strategy where agents start with cheaper models (CHEAP tier) and escalate to more expensive, capable models (CAPABLE -> PREMIUM) when needed. This approach balances cost efficiency with accuracy for complex analysis tasks.

Release preparation agents

The ReleasePrepTeam coordinates specialized agents that run in parallel:

• TestCoverageAgent — Executes pytest --cov and parses coverage reports
• DocumentationAgent — Validates docstring coverage, README currency, and CHANGELOG presence
• CodeQualityAgent — Runs ruff linting and checks type hints and complexity metrics

Each agent inherits from ReleaseAgent and returns structured results through ReleaseAgentResult, including success status, tier used, findings, and cost tracking.

State management

Agents support Redis-based state persistence through AgentStateStore, enabling recovery from failures and maintaining execution history via AgentExecutionRecord. The AgentRecoveryManager handles automatic retry logic with exponential backoff.

Framework adapters

The system provides adapters for popular AI frameworks through lazy-loaded functions:

• get_langchain_adapter() — LangChain integration
• get_langgraph_adapter() — LangGraph workflows
• get_autogen_adapter() — AutoGen multi-agent systems
• get_haystack_adapter() — Haystack NLP pipelines

The wrap_wizard() helper quickly converts wizard instances into agents, while the WizardAdapter provides native integration.

Quality gates and reporting

Release readiness is determined by configurable quality gates defined in QualityGate objects. The ReleaseReadinessReport aggregates all agent results into an approval decision with confidence metrics, blockers, warnings, and cost summaries.

Source files

• src/attune/agents/**
• src/attune/agent_factory/**

Tags: agents, ai, release