Level 3: Resumable Coding Loop

The most operationally demanding example in the repo applies the same loop lifecycle to code changes inside a sandboxed fixture repository.

Pattern

• Multi-agent fleet: observer, planner, and implementer agents divide inspection, action selection, and code generation.
• Sandboxed tools: file reads, writes, and shell execution stay confined to a copied fixture repository.
• Parallel evaluators: tests, requirements checks, and failure-memory guards all contribute to the verification decision.
• Failure memory: repeated approaches are recorded and blocked before the loop wastes more iterations.
• Resume support: the loop persists enough state to continue after interruption.

Architecture

Example configuration

The coding run gives itself more room, but it is still bounded:

run = LoopRun(
    example_id="level_3_coding_fleet",
    goal="Add expiration/TTL support to KVStore while preserving backward compatibility.",
    budgets=BudgetConfig(
        max_iterations=max_iterations or 6,
        max_model_calls=30,
        stagnation_threshold=3,
    ),
)

The orchestrator records failed approaches and blocks repeats after three tries:

def record_failure(self, approach_hash: str, error: str, iteration: int) -> None:
    for record in self.failure_history:
        if record.approach_hash != approach_hash:
            continue
        record.retry_count += 1
        if record.retry_count >= 3:
            record.blocked = True
        return

Verification mixes repository tests with domain checks:

def evaluate_requirements(state: LoopState) -> EvaluationResult:
    kvstore = orchestrator.fs.read("kvstore.py") if orchestrator.fs.exists("kvstore.py") else ""
    has_ttl = "ttl_seconds" in kvstore and "_is_expired" in kvstore and "time.time()" in kvstore

Why it matters

This example demonstrates that loop engineering is not just for text generation. The same controller can manage tool execution, requirement verification, repeated-failure avoidance, and resumable progress without handing control of the workflow to the model.

If you want the design principles behind these examples, move to the concepts section.