Loop compression

Early Meal Companion runs treat every week as novel. The agent generates candidates for every evening, assesses each, composes, critiques, patches, and sometimes regenerates accepted work. After dozens of accepted busy-week plans, the pattern stabilises. Tuesday and Thursday are low-time evenings. Spinach must move before midweek. Leftovers are acceptable once. Re-running full deliberation every week wastes money and introduces variance where the terrain is familiar.

Helpful context: Applicability conditions explain when a compressed strategy may run. How expertise forms explains where the strategy came from. Minimum sufficient loop defines when the agent may stop. This page is the building implication: compression as an explicit code path, not a hidden optimisation.

Guard pass, compress. Guard fail, expand.

Before reusing the Busy Week Pattern, the guard checks concrete conditions from the grounding reference. Are at least two low-time evenings available? Are busy evenings clearly identified? Is any unfamiliar medical restriction present? Are leftovers acceptable this week? Does the temporary request conflict with the strategy?

When the guard passes, the compressed path anchors the three busy evenings first, places spinach before Wednesday, uses one accepted leftover-based subplan, composes remaining evenings around variety, runs exact checks and one complete critique, and patches locally if a single weakness remains.

When the guard fails, the agent returns to full deliberation. Generate, assess, compose, critique. The compressed path does not pretend it still applies because it usually works. Hidden optimisation is how systems become untrustworthy. A team that skips steps when it is obvious has no evaluation record, no rollback, and no way to know whether Tuesday's failure came from compression or from a weak cognitive unit.

Incoming weekApplicability guardCompressed strategyExact checks + critiqueFallback to full loop if guard fails

Compression is a branch, not a silent shortcut.

Why compression must be explicit

Explicit compression means versioned expertise: the Busy Week Pattern v0.2 is published experience promoted to expertise with documented applicability, not a prompt tweak nobody owns. It means testable guards where the same inputs produce the same pass or fail every time, with guards belonging in code or deterministic checks rather than model discretion. It means evaluation on both paths, so suites cover compressed and full deliberation on the same cases before promotion. It means fallback always available: new medical restriction, unusual guest count, conflicting constraints all return to full loop without shame.

Premature discipline is the opposite failure: publishing compression before underlying judgments are validated. Compression before expertise is a hidden shortcut with extra steps. unit cost falls with history describes the economic win when compression is earned through evidence, not assumed on day one.

What this looks like in practice

Month one: seven evenings times five cognitive units times two loop iterations equals roughly seventy cognitive unit calls per plan. Full deliberation every time. Acceptable while learning.

Month six, Busy Week Pattern promoted and guard passing on familiar weeks: anchor three busy evenings, compose four, one critique, zero regeneration on accepted evenings. Roughly twenty calls. The plan arrives in seconds, not minutes. The cognitive units still exist. AssessMealPracticality still runs on busy evenings. GenerateCandidates still serves unfamiliar weeks. What changes is agent strategy when expertise licenses compression. Redundant candidate generation and full recomposition drop away when the week matches a validated pattern.

Consider the second originating request used for compression material:

Plan dinners for four. Monday, Tuesday, and Thursday are busy. Use the spinach before Wednesday. We are happy to eat leftovers once this week. Avoid pasta because we had it several times recently.

The guard must notice temporary pasta avoidance as context, not durable knowledge. It must confirm leftovers are acceptable. It must fail closed when an unfamiliar medical restriction appears even if the week looks busy-shaped. Each condition is testable. Each failure mode returns to full deliberation with logging so operators see terrain shift.

Building compression as a code path

Agent authors implement compression as a branch visible in strategy code and trajectory logs. The agent evaluates the applicability guard on interpreted week and working state. If the guard passes, it loads the named expertise pattern and executes the strategy skeleton. It runs exact checks and critique required by the pattern. If any check fails, it falls back to full loop and records guard or pattern failure reason. Pass and fail rates are instrumented so promotion and rollback decisions use data.

See maturity moves settled work out of runtime for where compression fits in maturity, and the agent loop for the full deliberation path compression may shorten.

Expertise promotion and demotion

Compression patterns promote from experience with evidence. Selected trajectories become experience records with context, cost, corrections, and evaluation. When busy-week plans succeed repeatedly under similar guards, the team extracts Busy Week Pattern v0.1 with applicability skeleton and failure modes. Suites compare compressed and full paths on the same cases. Promotion to v0.2 is a library event with an owner, not a prompt edit nobody remembers.

Demotion matters as much as promotion. When guard fail rates rise because household constraints changed, the pattern version rolls back or narrows applicability. Compression without a demotion path is how organisations keep using strategies that silently fail on new terrain.

Guard pass and fail rates alongside path selection belong in operational logs. Operators see terrain shifts before user complaints arrive. Compression rewards teams that capture experience during ordinary runs. Without experience records, every week stays novel and full deliberation remains the default.

Teams sometimes confuse compression with caching model outputs. Caching hides variance without guards, versions, or eval on both paths. Compression changes agent strategy under explicit expertise promotion rules. The distinction matters when debugging Tuesday failures.

Compare compressed and full paths on held-out weeks that almost match the pattern but fail one guard clause. Those edge cases prevent overfitting the Busy Week Pattern to one household's phrasing and keep guards honest as requests vary. Instrument both paths in staging before promotion. Side-by-side trajectory logs show whether compression skipped assessment on evenings that still required AssessMealPracticality or removed critique when material weaknesses would have surfaced. Staging surprises are cheaper than production surprises.

When promoting compression, the guard publishes alongside the pattern in the expertise library so other teams know when reuse is licensed. Hidden guards in agent code recreate the inconsistency behavioural standards are meant to prevent. Fallback behaviour belongs in the same specification section as the pattern so operators and engineers learn what happens when the guard fails on a busy week with a new medical restriction without reading agent strategy source.

Compression remains optional acceleration, not the default path, until metrics show guard pass rates stable across seasons or request mixes relevant to the domain.

Loop compression replaces full deliberation with expertise-backed strategy when applicability guards pass: explicit, eval-backed, reversible.

Introduction to Thoughtware . Ch. 26

Common mistakes

Compressing before expertise is validated. Patterns need promotion from experience with suites, not intuition. A pattern without evidence is a guess with fewer steps.

Non-testable guards. "Looks like a busy week" is not a guard. Guards belong in code with deterministic pass or fail on the same inputs every time.

Skipping eval on the compressed path. Faster wrong is still wrong. Suites cover both paths before promotion declares compression safe.

Never falling back. Compression that cannot expand becomes brittle silently. Guard failure must return to full deliberation with logging, not produce a degraded plan nobody debugs.

What to do next

Identifying loops that repeat with stable outcomes across runs is how the first compression candidate surfaces. Extracting an expertise pattern with explicit applicability guards, evaluating compressed and full paths on the same suite before promoting compression, and instrumenting guard pass and fail rates so terrain shifts surface quickly form the implementation path. Compression is a branch engineers can see, test, and roll back rather than optimism dressed as optimisation.

See applicability conditions, how expertise forms, and premature discipline.

Read next: Intelligence-native products.