Scale: multi-agent systems and feedback control

Hard rule: do not allow several workers to edit the same files simultaneously without ownership, locking or merge rules.

The moment work is split across several agents, you inherit the classic distributed-systems problems, including race conditions, conflicting edits, stale context and duplicated effort, but without the mature tooling that exists for conventional distributed systems. Throughput rarely scales with the number of agents, while the cost of merging and reconciling their work grows quickly. This is the pattern where impressive demos most often fail to survive real workloads.

A generate, evaluate and repair cycle is a feedback controller, and it fails the way controllers fail: oscillating between two solutions, overcorrecting so one fix breaks another criterion, settling into a local optimum, satisfying the grader without meeting the real goal, or simply never declaring completion. Each iteration costs money and elapsed time, so these failure modes are budget events, not curiosities. The controls are ordinary: a target, a cap, a budget and explicit stop conditions.

Long-running work must resume from explicit, stored state. When progress lives only in a growing transcript, cost rises with length, older decisions get crowded out and quietly forgotten, and an interruption loses work that has already been paid for. Checkpointed state makes a run resumable, auditable and cheap to continue, and it is the difference between a process that survives a deployment and one that has to start again.

Every decision should be routed to the cheapest layer that can carry it reliably: deterministic code first, then a small model, then a strong model, and a human only where judgement or accountability genuinely requires it. Sending everything to the largest available model is the most common reason a promising feature has per-request costs that cannot scale, and it also adds latency users notice. Routing is a design decision with a direct margin impact.