Agent Architecture Lens

The ReAct loop (reason, act, observe, repeat) is the atom of every agent and is correct until the task outgrows it. A loop is a graph with one node and one edge back to itself. Architecture follows the shape and duration of the task, never the elegance of the diagram.

Two leitwörter anchor every pass; state them by name as you review. A graph is not an upgrade: structure answers a requirement the task actually has, and without one the loop stays. The ledger, not the transcript: progress that must survive anything lives in a durable artifact, never in the message history the model hopes to remember.

The six signals

Before recommending structure, state which of the six signals fire, by number, with evidence:

1. State exceeds one context window.
2. The run must survive crashes: losing it costs minutes of work, real spend, or a waiting user.
3. Human approval is required mid-run, possibly hours later.
4. The workflow has parallel branches that are genuinely independent.
5. Multiple models or agents must coordinate, with different models, tools, or permissions per step.
6. Execution must be replayable and auditable after the fact.

Zero signals: "No signal fires; this stays a loop. A graph is not an upgrade." Stop there. Otherwise each fired signal names the mechanism the design owes, and only the checks for those mechanisms apply.