Intermediate · About 25 minutes with practice

Design loops that know when to stop

Turn observe–act–evaluate into a bounded, recoverable process with evidence-based improvement.

Make this yours

A loop contract with stop conditions, retry rules and a versioned improvement test.

Start with your own answer ↓
Speculative Academy artwork: intelligence that tends its own habitat.
Imagined world · Intelligence that tends its own habitat

A moment before the explanation

Why is a successful scheduler heartbeat different from a completed task?

Give yourself a first answer, even if it is incomplete. Keep it in mind as you read—or write it in your field note.

01

A loop needs a state you can inspect

A loop repeatedly observes a situation, chooses a permitted action and evaluates the result. Define the state explicitly: waiting for input, preparing, checking, done, blocked or failed. A timestamp showing that a process woke up does not establish that useful work completed.

Consider a weekly source digest. The trigger may be a schedule, but the work identity should name the digest period. Store the input set and the output artifact separately from the scheduler's status. A second invocation for the same period should find existing state rather than silently create another publication.

02

Recovery is part of the design

Use a stable work identifier to recognize repeated requests. If an external action times out, its outcome may be unknown: the service may have completed it while the response was lost. Blindly retrying can duplicate the effect. Inspect a receipt or use an idempotent operation before trying again.

Bound attempts, elapsed time and resource use. A failure should preserve enough context for diagnosis without retaining secrets. Some failures deserve a retry; a missing authorization or contradictory source usually deserves a stop. Decide that distinction before leaving a process unattended.

03

Learning requires a comparison

A loop that edits its own prompt has changed, but has not necessarily improved. Keep a fixed evaluation set and compare the candidate with the previous version. Preserve regressions and difficult cases. Promote a change only after a named criterion is met.

In this Academy, reflection and revision are learner-controlled. No hidden model trains on your notes. You can export a record of the initial attempt, feedback and transfer exercise. The same principle applies to an agent system: make the change and its evidence visible before calling it learning.

Explore a bounded agent loop

The same workflow. Three different outcomes.

  1. 1Observe
  2. 2Prepare
  3. 3Check
  4. 4Done
Step 1 of 4

Observe

Three public source files are available. The requested output is a local comparison draft.

An illustrative trace. No model, tool or external action is executed.

A worked example · Illustrative

A bounded digest loop

Work ID: digest-2026-W38
Trigger: manual invocation in this exercise
States: waiting → drafting → checking → done
Failure states: blocked | failed | outcome-unknown
Budget: 2 attempts, 10 minutes
Retry: only after the previous effect is known
Done: output exists AND every required check passes
Improvement: compare v2 against v1 on fixed fixtures

Put the idea to work

Your practice

  1. Draw the normal path and three failure paths.
  2. Name the work ID and the evidence required for done.
  3. Add a timeout case where an external effect is unknown.
  4. Change one instruction, then compare old and new behavior on the same examples.

Make the idea available when you need it

Your field note

Keep your first answer, make an artifact and try the idea somewhere new. Save only what you want stored on this device.

Check one distinction

A send request times out. The system does not know whether delivery happened. Which state is honest?

Take this practice into your own agent.

Copy a coaching brief, then choose your host in the Academy installer.

Choose a host ↗

Your text stays in this page until you save or export. Local saving can be unavailable in restricted browsers. Curriculum 2026-09-15.1. Import, export and browse your notebook ↗

Follow the evidence further

Primary sources & deeper study

Selected 15 September 2026. Provider docs govern current setup; older courses and engineering articles provide foundations.

Starlight Intelligence System · Versioned source code

Loop graph implementation

Inspect the implementation beside your own loop contract. This lesson does not execute SIS.

Starlight Intelligence System · Versioned source

Operational work graph

Event identity, proof gates, provenance and memory promotion. Read as an architecture contract.

Anthropic · Engineering article · 2024

Building effective agents

A useful foundation for choosing between fixed workflows and model-directed action.

Take the method into your work

Practice evaluating failure cases
Another question to follow

Make relationships precise enough to use