GrowthHarness GitHub

The mechanism

Pre-registered LearningThe self-evolution engine for AI agents

Fix the standard first. Then let the result speak.

Any agent that evolves at the context level is a memory system underneath. Here one memory is a triple of intent, experience and utility, laid out on the same timeline science uses to pre-register an experiment: state the hypothesis and the standard first, run the experiment, then compare. The result is never allowed to change what counted as success.

Intentwritten, then frozen
Experiencerewritten from evidence
Utilityearned by later use

The idea

Memory, borrowed from science.

In a lab, a pre-registered study publishes its hypothesis and its success criteria before the experiment runs, so the result cannot be explained after the fact. An agent that is supposed to get better with use needs the same discipline.

Pre-registered learning on one timelineIntent is written and frozen before the work. Experience is rewritten as evidence arrives and is scored against the frozen intent when the outcome returns. Utility stays empty until a later task adopts the experience, then moves by 20 percent toward each new score.IN SCIENCEhypothesis + standardrun the experimentcompare, reviewadjust the next runBEFOREDURINGAFTERLATER TASKSFROZEN01IntentWhat were we trying to achieve?Learning Agent writes it · frozen on hand-back02ExperienceWhat did the work teach us?Learning Agent · rewritten from evidence03UtilityHas it helped later work?Computed by code · no model writes itgoal · context · expectationhow it will be scoredlocked: nobody can change it from herecomparedeliveryreportfeedbackreportpublicationeach new piece of evidence rewrites the findingsoutcomescored against the frozen standardreward 0 to 1, or unknownempty until someone uses ita later taskadopts it, with a receiptits score flows backutility moves 20% of the waytoward each new score

Scroll sideways to see the whole diagram.

01Intent

What the practice is for, written before it starts: the goal, the conditions (known facts and constraints) and the expectation, meaning what a correct result looks like before anyone acts. It usually corresponds to the user’s request, decomposed into a proposition.

02Experience

A running record of what the practice really did: how it judged, what it did, what came of it and what was learned. It is written in three phases, before, during and after, and always cites its evidence.

03Utility

Whether the experience helped later work. It is estimated only from the results of later tasks that used it, never from the practice that produced it. It is empty until someone has used it.

Why memory

There are three broad ways for an agent to evolve. This engine takes the third.

  • Parametric. Turn experience into datasets and keep fine-tuning a model. The most interesting, and the hardest to land in a real business today.
  • Harness-level self-modification. Make the harness itself modular, so a learning agent can build plug-ins and load them into the runtime or the agent loop. Promising, and close to commercial use.
  • Memory-based. Give every agent a memory shaped by its own business, so the context itself evolves. Most projects can land it today, and its effect is widely underestimated.
Intentone per practice
Goal
The business purpose of this practice.
Context
Known facts and constraints.
Expectation
What a correct result should look like, set before acting.

Before

Fix the exam first.

Before any work starts, a Learning Agent writes the hypothesis, the evaluation agreement and the observation plan, then hands the page back. From that moment nobody can change it.

The evaluation agreement

  • Objective and basis. Which goal is being judged, and which business material it rests on.
  • Inputs and constants. The measurements it will use, with names and units that must match the data source exactly, and fixed numbers such as targets and thresholds.
  • Steps and output. One operation per step, from add, subtract, multiply, divide, take the smaller, take the larger. One step is marked as the final score, and it must land between 0 and 1.
  • Score meaning. What a high or low score stands for.
  • Unavailable when. The conditions under which the result cannot be judged at all.
  • Comparison key. A label for the scoring basis. A score can later flow back only to an experience that used the same key.
Evaluation agreement · v1numeric
Objective
Clicks on work-context guides
Inputs
clicks (count), impressions (count)
Constants
target = 120 clicks
Step 1
ratio = clicks ÷ target
Step 2
score = min(ratio, 1)
Output
Step 2
Unavailable
Any day of the window is missing
Comparison key
clicks-vs-target

Two kinds of agreement

Numeric: a real formula, calculated by code once the result arrives. The model takes no part. Review: a written rubric, frozen the same way, scored afterwards by the Learning Agent, who first judges whether the evidence meets the agreement.

The observation plan

The data source (Search Console, or “unavailable” with the reason), the start (always the publish day), the number of days (chosen by the agent for this case, never a fixed period in code), the signals (clicks, impressions, click-through rate, position) and why.

Learning never blocks the business

Until preparation finishes, the Leader cannot delegate and the system refuses. If preparation keeps failing it is marked skipped, and the Leader carries on with the gap in plain view.

Adopting an experience

  • Search finds candidates, nothing more. Semantic search returns experiences that look similar. Similar does not mean relevant, and relevant does not mean useful.
  • The Leader reads the evidence and decides. If it uses one, it declares so in the very same write that delegates the work. The two succeed together or not at all.
  • The receipt is permanent. It records which experience, which version, which decision it changed and what is still uncertain. No receipt, no use, and it cannot be added afterwards.
Adoption receiptwritten at delegation
Experience
“Titles answer the search question”, version 3
Changed
The titles of all three articles
Uncertain
It was only tested on guides, not comparisons

During

The system keeps the evidence.

The specialists never know learning exists. Every formal event leaves evidence that no agent had to write, and the Learning Agent turns that ledger into findings.

Evidence kindWritten byWhat it holds
reportA database triggerA child task completed, failed or is waiting. A reviewer’s comments arrive with its report.
deliveryThe same write as the submissionA topic submission, as a snapshot of that moment.
feedbackA trigger, or the LeaderA rejection or handover by a reviewer, an approval or return of a topic, or the Leader’s coordination judgment.
publicationA triggerAn article went live: time, address, language. Recorded even if the Leader has already finished.
outcomeThe system, or the LeaderThe measured or reviewed result, from Search Console or from evidence the Leader submits.
observation_gapThe systemWhy something cannot be observed, for example no plan or the domain is not live. It never pretends to watch.

Stamped by the system

Each piece of evidence carries tags put on at the moment it is created: which task, which parent task, which assignment, and which subjects, with their type, id and version. The model cannot change them. If the same evidence id is written twice with different content, the database refuses.

In strict order

Every new piece of evidence queues a request, and requests for one practice are handled strictly in order, so later evidence is never summarised before earlier evidence.

Findings

  • Rewritten as a whole. The Learning Agent reads the new evidence and the existing record, then writes the full set again. It does not append.
  • Scoped. About one object (an article, a topic group), about the whole practice, or unknown when the evidence has no owner.
  • Four kinds. Adjustment: a change that was really made, because advice alone does not count. Delivery: what was actually handed over. Conclusion: what was learned, with its uncertainty. Limitation: which evidence is missing and which claims are unsupported.
  • Always cited. Every finding lists at least one piece of evidence.
Finding · adjustmentabout article 2
Text
The title changed between the first and second version. The reviewer said it did not answer the search the page targets.
Evidence
The reviewer’s report, and the writer’s resubmission report
Scope
Object: article 2, version 2

After

Google marks it. Code does the sums.

Weeks later the real numbers come back and are set against the frozen agreement. The result evidence is never written by the Learning Agent: it has exactly two sources.

One observation window per articleThree articles published on different days each get their own 28-day window. The first is closed, collected and scored. The second closes tomorrow. The third is still collecting and its score stays unknown.Article 1published day 0Article 2published day 2Article 3published day 5day 0day 7day 14day 21day 28day 35todaywindow closed, collected, scoredcloses tomorrowstill collecting. Its score stays unknown.Every article is measured from its own publish day, in Search Console’s own time zone. The window is collected the day after it closes.

Scroll sideways to see the whole diagram.

How a window opens

When an article goes live a trigger records the publication evidence. Every minute the scheduler opens one window per language version, counted from that article’s own publish day for the planned number of days. Dates follow Search Console’s own time zone, Pacific time, so they never drift by a day. If there is no plan or the domain is not live, an observation gap is recorded instead.

When it is collected

The day after the window closes. Each day must have data, and Google must not flag it as not final. Anything missing or failed is collected again the next day, until it is complete. A day with no data is not zero, and nothing is filled in.

How pages are merged

Clicks and impressions add up across the practice’s pages. Click-through rate is total clicks over total impressions. Position is averaged weighted by impressions. The result counts as complete only when every delivered article has every window complete.

The two sources of results

  • Numeric agreements: collected by the system from Search Console. A new result is stored only if the measurement changed, and each change queues a new assessment, so a numeric outcome can arrive several times. First an incomplete one, with an empty score, then a complete one.
  • Review agreements: submitted by the Leader after delivery. The code insists it uses the frozen version unchanged, never replaces an automatic collection by hand, cites only this practice’s own raw evidence, and keeps the measured names and units from the agreement. The Leader may hand in facts and limits. It may not write a score.
Assessmentpt-BR · 3 articles
Reward
0.71 All three windows complete. Calculated by code from the frozen formula.
Or
empty. One window is still open. Never replaced by 0.
Forced empty
no agreement, incomplete outcome, a missing measurement or unit, a formula that cannot be computed (such as dividing by zero), a mismatched agreement version, or evidence the agreement does not accept.

Utility

What an experience is worth.

A practice that scored well is not proof that what it learned is right. It might be luck, or search might simply be rising that month. So an experience is not judged by how its own practice went, but by how the work that used it went afterwards.

  • Three fields, none written by a model. The score from 0 to 1, empty until some task has used it and been scored; the number of effective uses; and the algorithm version, today ema-0.2-v1.
  • A practice cannot use itself. It can only use experience left by other practices on the same site, and only when it declared the use at delegation. Merely searching or reading counts for nothing.
  • A comparison, not a vote. When a later practice is scored, the Learning Agent decides which of its adopted experiences were about the same thing under comparable conditions. That takes reading, so it is the model’s call.
  • Then the code checks three things. The experience was declared at delegation. The agent read its latest version, and an older one is sent back to be read again. And the scoring basis matches: both sides must carry the same comparison key.
An experience that scored 0.8, then 0.3, then 0.6 ends at a utility of 0.680.00.51.0Titles answer the search question · v3task scoreutilityutility 0.80Task A · first useutility 0.70score 0.3Task Butility 0.68score 0.6Task C20% of the gap

One nudge of 20%

Scores are ordered by when each practice started. The first score becomes the utility. Every later score moves it 20% of the way toward the new score. A first use of 0.8 gives 0.80. A second use scoring 0.3 gives 0.70, not 0.30. It takes several poor results in a row to move it far.

Rules for the rows

An empty score is not counted. Several agents in one practice using the experience count once. A practice that is re-scored replaces its row instead of adding a sample. And every change recomputes everything from scratch, so a corrected row always gives the right answer.

Why it is split

Utility is where a self-learning system most easily goes wrong. A model asked to judge usefulness is easily persuaded by a handsome result, and a practice’s own score used as its experience’s score is marking its own homework. So understanding stays with the model, and rules that must not be bypassed stay with code.

Utility does not decide for the Leader. It only shows how an experience did in later work. Whether to use it for this task is still the Leader’s call, and utility is estimated from later use, not proof of cause.

Who does what

One practice, start to finish.

Suppose the Leader wants three English articles on “how to build English reading skills every day”, aiming at search traffic. This is the whole journey, and who touches what.

One practice across the actors and the timeTwelve numbered steps across six lanes: the Leader starts a practice, triggers queue a prepare request, the scheduler starts a Learning Agent who writes and freezes the standard, the Leader adopts a past experience and delegates, specialists work, triggers record evidence, the Learning Agent rewrites findings, articles publish, the scheduler collects Search Console data, the Learning Agent scores the result, and code recomputes the utility of the adopted experience.BEFOREthe standard is written, then frozenDURINGwork, evidence, findingsAFTERthe result arrives and is scoredLeaderSpecialistsDatabase triggersSchedulerLearning AgentSearch ConsoleFROZEN123456789101112

Scroll sideways to see the whole diagram.

  1. The Leader starts the practice. It judges the work worth learning from and hands in the goal and what is known.
  2. The system queues a prepare request. A new practice record is created in the database.
  3. The scheduler starts a Learning Agent. It is not resident. Each request starts a new task that reads the record and the evidence, submits one result and ends.
  4. The standard is written, then frozen. Intent, hypothesis (“articles with concrete exercises earn more clicks than general advice”), agreement and observation plan. Nothing changes after this.
  5. The Leader adopts an experience. It searches, reads, decides to use “titles answer the search question”, and declares so in the same write as the delegation.
  6. Specialists do their work. They only follow the Leader’s brief and know nothing of learning.
  7. Triggers record evidence. Reports, deliveries, feedback and publication, with no agent reporting anything.
  8. Findings are rewritten. Article 2 is sent back once and its title changes, so the agent writes that adjustment, citing both reports.
  9. Articles go live. For each language version a window opens from its own publish day. Then nothing runs for weeks except the scheduler.
  10. Windows close and data is collected. The scheduler pulls Search Console day by day, merges the pages, and stores the result if it changed.
  11. The result is scored. A new Learning Agent compares it with the frozen agreement. Code computes a numeric score, and if one article’s window is still open the score stays empty until it is complete.
  12. Utility is updated. In the same write, code recomputes the worth of the experience adopted in step 5, but only if the two are comparable.

The Leader

The only business role that touches learning. It decides whether to open a practice, which experience to use, and when to submit results.

The specialists

They know nothing about learning, on purpose. Keeping one entry point keeps information consistent and the record in one place, which is what makes an experience trustworthy.

The machinery

Database triggers remember which practice a task belongs to and queue requests. A per-minute scheduler handles requests, collects Search Console data and refreshes the semantic index, which recomputes a vector each time a record changes so the Leader can find the right experience.

The standard is fixed before acting. A score only flows to the experience that was really used.

Results can be compared with the standard, never re-explained. That is what lets an agent get better without ever grading its own homework.

How it keeps running The Agent team Read the source