Writing

AI agents do not remove systems engineering

Tool calling is a new I/O surface. Queues, identity, permissions, and failure handling remain the job.

2026-08-20notes

An agent that can call tools is a distributed system with a probabilistic control plane.

That is not a metaphor. Tool calls have latency, partial failure, retries, and authorization. The model is allowed to choose the next call; it is not allowed to invent the semantics of your CRM, your queue, or your database.

What still has to exist

  • Explicit contracts for every tool: inputs, outputs, idempotency keys, error classes.
  • A place to put work that cannot finish in one request.
  • Permissions that are not “the prompt said it was okay.”
  • Evaluation that includes tool traces, not only the final paragraph.

The systems work people skip

Orchestration graphs that hide the data flow. Memory that cannot be queried. Retries that double-charge or double-write. Logs that contain secrets because the fastest path was print(payload).

Agents make those omissions louder. They do not make them optional.