Beyond ReAct: Dynamic Plan Revision in Autonomous Coding Agents

The canonical ReAct (Reason + Act) pattern revolutionized tool-using LLMs by interleaving internal thought tokens with tool invocations. However, when applied to long-horizon software engineering tasks—such as refactoring an entire microservice or migrating database schemas—standard ReAct suffers from catastrophic failure modes:

• Greedy Error Compounding: If an agent misunderstands an API schema on step 3, subsequent steps build upon flawed assumptions until the context window explodes.
• Context Window Exhaustion: Accumulating raw command output, stack traces, and verbose build logs pushes critical architectural constraints out of active attention.
• Plan Drift: Without structured checkpoints, agents deviate from the user's primary specification to fix tangential formatting or lint warnings.

The Dynamic Plan Revision Architecture

Modern agent harnesses replace linear ReAct loops with hierarchical plan execution graphs featuring explicit rollback and branch verification.

stateDiagram-v2
    [*] --> PlanSynthesis: User Specification
    PlanSynthesis --> ExecutionStep: Approved Plan DAG
    ExecutionStep --> ToolExecution: Execute Action
    ToolExecution --> VerificationCheck: Collect AST & Test Output
    VerificationCheck --> ExecutionStep: Pass (Next Step)
    VerificationCheck --> PlanRevision: Fail / Constraint Violation
    PlanRevision --> ExecutionStep: Updated DAG / Rollback
    ExecutionStep --> Completion: All Steps Verified
    Completion --> [*]

1. Structured Step Contracts

Rather than free-form natural language thoughts, the agent generates and commits to structured JSON step definitions:

{
  "stepId": "migrate-auth-tokens",
  "dependencies": ["schema-migration-applied"],
  "intendedModification": {
    "files": ["src/auth.ts", "test/auth.test.ts"],
    "invariants": [
      "Zero breaking changes to verifyCredential signature",
      "All unit tests in auth.test.ts must exit code 0"
    ]
  },
  "rollbackPlan": {
    "gitCheckout": "HEAD~1",
    "discardPaths": ["src/auth.ts"]
  }
}

2. Compacting the Observation Stream

Verbose compiler and test outputs should never be injected verbatim into the root planner prompt. High-efficiency agent runtimes pass raw stdout through a Summarization Filter:

• Diff Extraction: Extract only the lines surrounding compile errors or test failures (context $\pm 3$ lines).
• AST Scoping: If a refactoring tool touches 40 files, present a summary table of modified exports rather than thousands of lines of syntax trees.

3. Empirical Results

In comparative benchmarks evaluating autonomous bug-fix pipelines on 150 complex repository pull requests:

By enforcing programmatic verification at each transition point, autonomous agents transition from unpredictable toy demos into resilient engineering assistants.