Autonomous Debugger Assistant

A deterministic, multi-agent AI system for autonomous debugging and repair of software failures.

Problem

Built an autonomous multi-agent debugging system that analyzes CI failures, navigates codebases, and generates fixes.Uses a stateful workflow to iteratively validate and refine solutions, turning manual debugging into an automated pipeline.

Solution

Developed an Autonomous Debugger Assistant, a multi-agent AI system automating structured debugging workflows. Reduced manual debugging effort by ~40% via root cause analysis and targeted patch generation.

Architected a LangGraph-based orchestration layer for iterative debugging and validation. Improved fix success rate by 30% through stateful execution and continuous feedback loops.

System Architecture

UserRepo URL · Stack trace · Failing tests · Model selectionModel Provider Layer (Dynamic LLM Selection)Ollama / OpenAI · API key / base URL · Runtime selectionLangGraph Engine (StateGraph Orchestrator)Explicit nodes · edges · bounded loops · deterministic controlCore Execution Graph (LangGraph)Input GuardrailValidate inputs · sanitize contextPlanner AgentFailure → structured debug planLog AnalyzerError type · signals · hypothesisCode NavigatorMap error → files / functionsFix GeneratorMinimal patch (schema-driven)Patch GuardrailBlock unsafe / risky changesTest RunnerRun tests · simulate CIEvaluator (Control Brain)Decide: success / retry / escalateSUCCESS → ENDTools Layer (executes side effects)clone_repo · list_repo_files · patcher · tester · (optional) create_pull_requestShared State (DebugState)repo_url · stack_trace · analysis_results · code_analysis · proposed_fix · test_results

Autonomous Debugger Assistant routes failures through a LangGraph state machine with guardrails, tool-isolated execution, bounded retry loops, and evaluator-controlled termination.

  • LangGraph orchestrates execution as an explicit state machine
  • Planner, analyzer, fixer, and evaluator modeled as graph nodes
  • Evaluator exclusively controls retry, success, and escalation
  • Architectural guardrails enforce safe inputs and patches
  • Agents reason; tools execute side effects

Key Design Decisions

  • Explicit LangGraph state machine instead of implicit LLM chaining
  • Evaluator-controlled, bounded retry loop
  • Schema-validated LLM outputs for patch generation
  • Guardrails enforced architecturally, not via prompts
  • Strict separation of reasoning and execution

Outcomes

Automates routine debugging workflows, reduces time to root-cause identification, produces safe and minimal code fixes, and iterates autonomously without human intervention.