Skip to main content

​
Iterations and Verification

AgentGate uses an iterative process with multi-level verification to ensure generated code meets quality standards. This page explains how iterations work and what verification levels mean.

​
The Iteration Process

When a run executes, it goes through multiple iterations:
1

Code Generation

The AI agent analyzes the task and generates code changes.
2

Verification

Automated verification checks run against the generated code.
3

Feedback Loop

If verification fails, results feed back to the AI for improvement.
4

Repeat or Complete

Process repeats until verification passes or max iterations reached.

​
Why Multiple Iterations?

AI-generated code doesn’t always work on the first try. The iteration process:
  • Catches errors early through verification
  • Provides feedback for the AI to improve
  • Ensures quality through multiple checks
  • Converges on working solutions

​
Verification Levels

AgentGate uses four verification levels, each progressively more thorough:

​
L0: Contract Verification

What it checks:
  • Syntax correctness
  • Type checking (for typed languages)
  • Import resolution
  • Basic linting
When it runs: Every iteration Example failures:
  • Syntax errors
  • Missing imports
  • Type mismatches

​
L1: Test Verification

What it checks:
  • Unit test execution
  • Test pass/fail status
  • Code coverage thresholds
When it runs: After L0 passes Example failures:
  • Failing unit tests
  • Below coverage threshold
  • Test runtime errors

​
L2: Behavioral Verification

What it checks:
  • Integration test execution
  • API contract compliance
  • Cross-component interaction
When it runs: After L1 passes Example failures:
  • Integration test failures
  • API response mismatches
  • Database interaction errors

​
L3: Sanity Verification

What it checks:
  • End-to-end scenarios
  • Performance thresholds
  • Security scans
  • Final quality checks
When it runs: After L2 passes Example failures:
  • E2E test failures
  • Performance degradation
  • Security vulnerabilities

​
Gate Plans

Gate plans configure which verification levels apply:

​
Default Gate Plan

By default, all relevant verification levels run:
  1. L0 always runs
  2. L1 runs if tests exist
  3. L2 runs if integration tests exist
  4. L3 runs for final validation

​
Custom Gate Plans

Configure verification for your needs: 
{
  "taskPrompt": "...",
  "workspaceSource": { "..." },
  "gatePlan": {
    "levels": ["L0", "L1"],
    "customChecks": {
      "lint": true,
      "typecheck": true,
      "test": true
    }
  }
}

​
When to Customize

Use only L0 for rapid prototyping where speed matters more than comprehensive testing.
Use all levels (L0-L3) for production-critical code where quality is paramount.
Emphasize L1/L2 when the codebase has good test coverage you want to leverage.

​
Convergence

A run converges when verification passes at all required levels: 
Iteration 1: L0 ✓, L1 ✗ (test failures)
Iteration 2: L0 ✓, L1 ✓, L2 ✗ (integration issue)
Iteration 3: L0 ✓, L1 ✓, L2 ✓, L3 ✓ → CONVERGED

​
Factors Affecting Convergence

FactorImpact
Task clarityClearer prompts converge faster
Codebase qualityWell-structured code helps AI
Test coverageMore tests provide better feedback
Task complexityComplex tasks need more iterations

​
Iteration Limits

​
Default Limit

The default maximum is 10 iterations.

​
Configuring Limits

Set custom limits per work order: 
{
  "taskPrompt": "...",
  "workspaceSource": { "..." },
  "maxIterations": 5
}

​
When Limits Are Reached

If max iterations reached without convergence:
  • Run status becomes failed
  • Best partial result may be available
  • You’re charged for completed iterations
If runs frequently hit limits, consider: improving task prompts, reducing task scope, or increasing limits for complex tasks.

​
Cost Implications

Each iteration incurs costs:
  • More iterations = higher cost
  • Verification level affects iteration cost
  • Failed iterations still cost credits

​
Cost Optimization

  • Write clear, specific task prompts
  • Use appropriate verification levels
  • Start with lower iteration limits and increase if needed
  • Break complex tasks into smaller work orders

​
Monitoring Iterations

Track iteration progress in run details: 
{
  "id": "run_abc123",
  "status": "running",
  "iterations": 3,
  "currentIteration": {
    "number": 3,
    "phase": "verification",
    "level": "L1"
  }
}

Runs

Learn about run lifecycle

Best Practices

Optimize for faster convergence