Agentic Variation Operators: Evolutionary Search in Local Sandboxes

Coding agents frequently get stuck in stagnation loops: an agent edits a file, runs a test suite, observes a failure, and tries the exact same broken fix repeatedly until it exhausts its token budget.

Agentic Variation Operators (AVO) introduce an evolutionary search mechanism into agent execution loops:

1. Level 1 Git Sandboxes: Sub-millisecond snapshot and rollback mechanics. If an agent edit breaks tests or increases regression counts, the sandbox immediately rolls back to the last known working state.
2. Level 4 Meta-Controllers: Stagnation detectors that analyze edit trajectories. If an agent loops on the same error twice, the supervisor intervenes with architectural guidance or forces exploration.

Architecture and Flow

Agent Edit Proposal ──> Execute Tests ──> Pass? ──(Yes)──> Commit Snapshot
                              │
                            (No)
                              ▼
                Rollback via Git Sandbox (0.8ms)
                              ▼
                 Supervisor Stagnation Check
                              ▼
                Intervene if Loop Detected

Key Takeaways

• Fast rollback mechanisms prevent error compounding during iterative coding.
• Meta-supervisors prevent infinite loops on difficult edge cases.
• Local isolation allows rapid parallel hypothesis evaluation.