Jump to text
On this page
Copy and source
Published source confirmed at last check

Source changed 2026-11-22 15:54:17 UTC · snapshot created 2026-11-22 15:55:11 UTC · last check 2026-11-22 16:35:10 UTC

Library.Reference:6.5.3.6 - AI: action, retained experience and model training

In ReAct, an observation returned by external action informs further reasoning and choice in the task. In Reflexion, evaluation of an attempt and reflection produce textual memory for later attempts. Retention matters to future behaviour, but recording context does not alone establish another control level. For a question about organizing adaptation, explain how retained conclusions are selected, used and checked. For a question about answering a request, one controller with memory may be an adequate description. Actor, evaluator and reflector name functional contributions, not necessarily distinct physical participants or levels.

In SCoRe, 2025, training changes model parameters to improve subsequent self-correction. This connects training work to later performance. A correct answer available during training is not thereby available during correction at test time. Preserve that difference when deciding what checking means the operating arrangement can use.

These papers support particular mechanisms and bounded experimental conclusions. E.23:11 explains further agent and harness mechanisms and their evidence limits. Relative advantage, transfer to another task and coverage of current AI practice require the corresponding comparison.

Choose what must change in the human–AI work. Start from the result its recipient needs. To diagnose a past attempt, establish what information reached the performer, which action they chose and executed, and how they used its return. To design a check before the first attempt, identify the choice the performer must make, what information will be available then, and what would let the recipient distinguish that choice from following a supplied answer. A compiler diagnostic, a current-state observation, an expert judgement and a preference supply different grounds. Relate each to the question it can settle. A proposed call needs an available interface and authority; its reported success needs the evidence required by the receiving use. Use SYSE.42 to bind and perform the operation and apply its qualified return.

Use the available grounds to decide whether a change is needed and what it should affect: the current answer, an input, retained experience, the procedure or assistance rule, model parameters, a person’s ability, or the provision of work. These are possible changed objects, not phases which every attempt must traverse. Explain how the chosen change will reach a later action and which observation would justify retaining it. Keep three decisions separate: ending the current search, accepting the current result and retaining a way for future use. A precise report of a missing fact can justify stopping without completing the requested intervention; a correct result on this occasion can suffice without developing a reusable procedure.

Retained experience needs both content and a way to use it. ACE v3, 2026, §§3 and 5 develops context through generator attempts, reflection and a curator’s incremental updates to identifiable entries. Later attempts consume the retained guidance; entries can be revised or removed. Its limits include dependence on feedback and reflection quality, and cases where a simple reusable rule is enough. Agentic Experience Learning, EMNLP 2026, §3 instead distinguishes records of episodes from policies for retrieving experience: a fast selection process updates its evidence about available policies, while a slower synthesis can propose new retrieval policies. Its reported financial comparison freezes learning before final evaluation. That experiment therefore does not establish the behavior of indefinitely changing deployed memory.

For the receiving work, determine whether the needed record was preserved, retrieved and used. SYSE.43 develops that connection. If a relevant return was dropped from the next input, restore it; if the procedure ignores an available result, change the consuming operation. Adding records or another reflector would address a different object. A fixed rule can update memory or select an action without another controller. When a participant selects and changes the rule itself, recover that participant’s observations, criterion, authority and subsequent check before describing the additional control relation.

A changed instruction must change performed work. GEPA v2, 2026, §§3–4 executes candidate prompts within a retained program, uses traces and task feedback to propose revisions, and compares candidates on their execution. This changes prompts while the underlying model weights stay fixed. Its candidate pool preserves alternatives useful on different cases; the data used to choose among them serve selection, while separate test data support a further-use claim. Apply C.40:4.7 when choosing what to retain from such search. A directly understood procedure repair can be sufficient; a prompt search is useful when its exploration and comparison supply a needed contribution.

Model training changes another part of the arrangement. TSRL, EMNLP 2026, §4 combines reinforcement learning with expert guidance selected for difficult states and changes the weight of that supervision during training. The returned product is an updated policy, whose later performance still needs checking. To construct such a change, SYSE.45 connects qualified interaction experience to the selected parameter update and tests new behavior alongside earlier useful behavior. Changing facts, execution access and permissions remain runtime contributions wherever the task needs them. Without appropriate training access or a qualified target, select another available change or return that missing condition.

Choose the human contribution by its result. A person can supply a fact unavailable to the model, clarify an intention, challenge an inference, make an authorized commitment, perform an action, or learn to judge and act. Recover the contribution needed now and the conditions for supplying it. Subject competence and authority to change a promise are different grounds. An approval request is useful only when its recipient can make the required decision; explanation or a corrected factual premise can be the missing contribution instead. Library.Reference:9.3 connects supported performance, development and joint work.

When the intended result includes a person’s ability, preserve that person’s needed judgement or action and observe it under the intended help. HCD.7 and HCD.9–HCD.13 supply the support, practice and evidence choices. Shen and Tamkin’s 2026 study makes the distinction consequential: in its short experiment with programmers learning an unfamiliar Python library, the group offered AI assistance scored lower on the subsequent unaided quiz. The main time comparison was inconclusive, and the study did not measure long-term retention. Its exploratory interaction patterns do not establish one generally superior teaching method. A good joint artifact and the learner’s acquired contribution therefore remain separate questions.

Reconsider the surrounding arrangement when its conditions change. A new model, tool or workload can make an old support unnecessary or reveal a missing one. Anthropic’s March 2026 engineering account describes removing some resets and subdivisions as models changed while retaining useful evaluation. Treat it as an experienced design account, not a controlled comparison proving one agent organization best. Use SYSE.46 to compare the actual arrangements at the intended reliance, including their needed checks and full burden.

Engineering.Reference:19.7 develops the connected choice from a failed attempt to a supported repair, construction and later use. It distinguishes a missing premise from a defective help decision, a lost input and an ignored return, and compares fixed-model changes with training. Use it when an actual attempt leaves the repair undecided. For a check commissioned before the first run, §19.7.5 shows how to construct two distinguishable outcomes from the available instructions or an executable model and how to leave the choice under examination to the performer. If the current method and its available support already obtain the required result, complete the work with them. Select further development for a specified improvement whose expected value warrants its burden.