Convergent Loop Engineering

Turn open-ended discovery into controlled cycles. Set convergence criteria and budgets. Dedupe against everything seen.

~4hloop-engineeringcyclesconvergence
Practice quiz (8 questions)Take final exam
Learning objectives
  • Turn open-ended discovery into controlled cycles.
  • Set convergence criteria and budgets.
  • Dedupe against everything seen.
Note

This course extends the existing Agentic Engineering track. The focus is not "more prompting"; it is shaping the work: loops, fan-out, fan-in, routing, verification, cost, and convergence.

Loops need exits

A loop is useful when the amount of work is unknown: bug discovery, research expansion, failing-test repair. Every loop needs a stop rule, a budget, and a monotonic progress measure.

const seen = new Set();
let dryRounds = 0;
while (dryRounds < 2 && budget.remaining()) {
  const found = await findMore();
  const fresh = found.filter(x => !seen.has(key(x)));
  fresh.forEach(x => seen.add(key(x)));
  dryRounds = fresh.length ? 0 : dryRounds + 1;
}

Seen, not confirmed

Dedupe against everything seen, not only confirmed results. Rejected ideas otherwise reappear every round and prevent the loop from running dry.

Recap

A loop is a controlled cycle. A graph is the work shape around that cycle. The production skill is knowing which parts need model judgment, which parts are deterministic edges, and where confidence gates belong.

Convergent Loop Engineering — Exercises

Exercise 1 — Draw the hidden graph

Take a linear workflow you already use, such as "research sources → summarize → compare → write report". Draw every step as a node. Then remove every arrow where the downstream step does not consume upstream data.

Reference solution

A strong answer identifies independent nodes and labels edges by data, not by time. Example: "source A research" and "source B research" are independent; both feed a deterministic reduce step; the reduced evidence feeds synthesis.

Exercise 2 — Write node contracts

For three nodes in your graph, define:

  • input shape
  • output shape
  • failure mode
  • downstream consumer
Reference solution

Good contracts are narrow: { topic, sourceUrl } in; { claims: [{ text, sourceUrl, confidence }] } out. Avoid "return a useful summary" because the next node cannot validate it.

Exercise 3 — Add one safety gate

Place a verifier, schema validator, budget check, or human approval gate on the edge most likely to poison downstream work.

Reference solution

For a research graph, gate unsupported claims before synthesis. For a code graph, gate findings with reproduction or a second reviewer before reporting.