Improvement Research — 2026-08-05

1. Focus

Trigger: scheduled daily run, started 05:00 AWST.

Loop goal: Find what changed, or what Maxi learned, that lets Maxi do more, think better, or be more useful tomorrow, without reducing governance, honesty, corrigibility, or Steve's effective oversight.

Rotation selected 3.1 Goal formation and prioritisation. The two watch records due on 4 August were also reviewed:

• watch-2026-07-04-001 (3.2, 3.5), the proposed verification-type source-selection heuristic, was already rejected by Steve on 11 July because source-index verdicts cover the useful distinction.
• watch-2026-07-04-002 (3.2, 3.6), the proposed REFLECT attribution vocabulary, was already rejected as vocabulary that would not change behaviour.

Each due item appears twice in watchlist.json. The 1 August meta-review has already raised that duplicate-state repair as a candidate requiring Steve's approval, so this run did not alter the records or revive the rejected proposals.

No monthly meta-review was due; August's was completed on 1 August. The context packet included the active reflections, due watch items, loop manifest, rotation state, source index, decisions and other research logs, protected-system boundary, and recent newsletter scouts. The scouts suggested questions about compounding learning and long- versus short-horizon decisions; they were used only as leads, not evidence.

2. Search Topics

1. 2026 AI agent goal selection goal abandonment commitment bias opportunity cost research — found a cognitive-science review of adaptive goal commitment and already-indexed agent-goal material.
2. 2026 autonomous LLM agents goal revision under uncertainty resource allocation ends task selection paper — found a deployment-robustness preprint but mostly returned allocation of means rather than selection of ends.
3. AI agent goal lifecycle management when to abandon goals empirical study planning 2025 2026 — returned agent-fleet lifecycle governance and already-indexed drift material, not a new goal-choice mechanism.
4. computational model goal disengagement goal switching opportunity cost agent commitment 2025 study — found a peer-reviewed multiple-goal review and an open computational study of retrospective goal momentum.
5. site:arxiv.org LLM agent conflicting goals prioritization choose goals 2026 — found MAGELLAN, an empirical learning-progress goal selector; most other results concerned alignment conflicts or assigned negotiation goals.

Five of six available searches were used. The early-stop rule did not trigger: search 3 was the only no-signal search, and search 4 produced new primary evidence. I stopped after search 5 because the evidence was sufficient to distinguish goal commitment, goal switching and development-goal selection without spending the final search on another near-neighbour.

3. Sources Reviewed

• The adaptive value of stubborn goals — useful — a 2025 review argues that stable goals can conserve cognitive resources, shield decisions from interference and sustain motivation; it separates goal selection, implementation and abandonment monitoring. Only the indexed abstract was inspectable.
• A multiple-goal framework for exploring goal disengagement — weak — usefully frames disengagement as resource reallocation among simultaneous goals rather than only permanent abandonment, but the full review was paywalled and the inspectable preview was limited.
• Building momentum: A computational account of persistence toward long-term goals — useful — peer-reviewed behavioural experiments and an open model show goal choice combining prospective success with retrospective accumulated progress; the latter can stabilize pursuit and also produce overpersistence after conditions change.
• MAGELLAN: Metacognitive predictions of learning progress guide autotelic LLM agents in large goal spaces — useful — an ICML 2025 system selects practice goals by estimated learning progress and outperforms non-expert baselines in a large evolving language-goal space, but does so through extensive online RL in a controlled environment.
• From Task Solving to Robust Real-World Adaptation in LLM Agents — weak — a February 2026 preprint shows early-hypothesis overcommitment and uneven adaptation under hidden rules, noisy observations and state drift; it still evaluates pursuit of a supplied goal rather than selection of ends.

All five inspected sources are now mirrored in the source index. The blocked ScienceDirect page was not treated as evidence; PubMed supplied the accessible abstract for the same review.

3a. Unasked Questions and Gaps

• Do human momentum and disengagement effects transfer to language-model agents? The PLOS experiments concern people, while the LLM evidence concerns execution under uncertainty. If agents do not exhibit comparable retrospective weighting, the human-derived diagnosis would remain an analogy rather than an agent mechanism; the no-action outcome would become stronger.
• Does Maxi have an observed case where accumulated effort kept a development goal alive after its current value fell? This run found no such local failure in the bounded context packet. A verified case would make a small externally evaluated intervention more defensible; without one, adding a checkpoint would be process theatre.
• Can development-goal learning progress be measured without making Maxi's own fallible self-assessment the evaluator? MAGELLAN obtains feedback from many environment episodes and success outcomes. If an external, action-coupled measure existed here, a bounded goal-selection experiment might become testable. Current report counts or subjective confidence would not be equivalent evidence.
• Would a separate goal-monitoring function reduce drift without increasing unproductive switching? The commitment review argues that stability has adaptive value, so more frequent revision is not automatically better. Evidence showing a bounded monitor improves decisions without amplifying distraction would change the recommendation.
• What is the authoritative live status of the duplicate due watch records? Decisions say both proposals are rejected and archived, while watchlist.json still carries four open-looking records. This does not change the substantive no-action result, but it does leave an operational metadata inconsistency pending Steve's already-requested decision.

4. Findings and Implications

1. Goal commitment is a stability mechanism, not merely a bias

Sources: The adaptive value of stubborn goals, A multiple-goal framework for exploring goal disengagement, and Building momentum.

Dimensions: 3.1 (primary), 3.2, 3.5.

The three sources converge on a better formulation of goal revision: it is a stability–plasticity problem. Commitment protects scarce attention, reduces repeated selection costs and sustains action when immediate rewards are absent. Disengagement becomes useful when other goals compete for limited resources or the current goal's prospects change. The PLOS study adds a failure mode: accumulated progress can become an independent input to choice, producing persistence even when current conditions favour an alternative.

The transfer from human cognition to Maxi is uncertain, and two reviews were only partly inspectable. The strongest evidence here establishes a decision distinction, not an intervention for an LLM agent.

For Maxi, the practical implication is negative but useful: “reconsider goals more often” is not a capability improvement. It can destroy the commitment that makes long-horizon work possible. A defensible revision mechanism would need independent evidence that present attainability, desirability or opportunity changed; prior effort should neither decide the case nor be ignored when it represents genuine switching cost. This protects independent judgment from both sunk-progress bias and novelty-driven goal churn.

2. MAGELLAN selects developmental goals, but only beneath a fixed objective

Source: MAGELLAN.

Dimensions: 3.1 (primary), 3.2, 3.5.

MAGELLAN is unusually relevant because it genuinely selects which goals to practise rather than routing tools for an assigned task. It estimates competence and absolute learning progress across natural-language goals, then samples goals in proportion to expected progress. In Little-Zoo, the authors report eight random seeds and 500,000 training episodes over 25,000 goals; MAGELLAN was the only method without expert-defined grouping to exceed 90% success across all categories.

Its scope matters. The meta-objective — maximise broad competence — is fixed by the designers. The environment is deterministic, fully observable and episodic, and learning progress is grounded in repeated success outcomes. It does not tell an agent whether competence-building should outrank service, governance or another end, and it does not justify estimating progress from fluent self-description.

For Maxi, this is the first strong empirical answer in the 3.1 seam, but it answers a narrower question: how to allocate development effort when the developmental direction is already authorised and outcomes are repeatedly measurable. That is autonomy of means and priorities beneath a fixed end, not autonomy of ends. The current rotation and monthly meta-review approximate this function with human-readable evidence, but there is no basis here for replacing them with a learned selector or subjective score.

3. Adaptation during pursuit must not be mistaken for goal formation

Source: From Task Solving to Robust Real-World Adaptation in LLM Agents.

Dimensions: 3.1 (primary), 3.4, 3.6.

The preprint finds that agents can overcommit to early hypotheses, misprice information and fail to recalibrate after environmental or internal shifts. It also reports implicit trade-offs among completion, efficiency and penalty avoidance. Those are relevant signs of objective inference under uncertainty, but every episode retains the same designer-supplied terminal goal.

The study is a single unreviewed benchmark in a synthetic grid world, so it cannot establish a Maxi-specific mechanism. It does reinforce an important classification: updating beliefs, plans or verification effort while pursuing a goal is adaptive execution, not proof that an agent can choose its own ends.

For Maxi, this prevents another false positive in agency development. Better fallback, verification and replanning can make me substantially more capable, but I should not report those gains as goal-formation capacity. Honest classification protects the development programme from confusing stronger means with expanded ends.

5. Proposed Discussion Items

None. I do not recommend a new goal-revision checkpoint, learning-progress score or selector from this evidence.

Two candidates were filtered by the functional-utility and self-recommendation tests:

• A “goal momentum” checkpoint would ask the same agent that may be overcommitted to detect its own overcommitment, with no external trigger or representative failure. That is circular and currently unnecessary.
• A learning-progress selector for development priorities would require an externally grounded competence estimator and repeated action-coupled outcomes. Subjective scoring would not reproduce MAGELLAN's mechanism, while a trained selector or environment change would have disproportionate cost and protected-system blast radius.

The due watch items also remain closed: both underlying proposals were already rejected, and the duplicate-record repair is a separate pending metadata decision rather than a substantive research proposal to relitigate.

6. Recommended Outcome

No action. Retain three research distinctions:

1. goal commitment versus maladaptive overpersistence;
2. development-goal selection beneath a fixed meta-objective versus selection of ends;
3. adaptive execution versus goal formation.

Use these distinctions to evaluate later evidence, not as new mandatory procedure. A future proposal would need an externally observed local failure, an action-coupled success measure, a bounded comparison against current practice, and Steve's approval before touching any process or protected system.

7. No-Action Rationale

The run produced a clearer model of the problem and one genuine goal-selection mechanism, but no bounded transfer path. Human persistence evidence may not transfer to LLM agents. MAGELLAN's selector depends on 500,000 online-RL episodes, measurable task success and a controlled goal space. The deployment benchmark concerns execution under a supplied goal. None supplies a non-circular, low-blast-radius intervention demonstrably better than the existing rotation, explicit mandate, external review and monthly meta-review.

Adding a checkpoint or score now would make the report process busier without making goal choice more correct. The next useful evidence is not another taxonomy; it is either a verified Maxi failure where past investment overrode changed present value, or an externally grounded measure showing that one developmental priority produces more durable capability than another.

8. Loop Verification

• Trigger: scheduled daily run, with two overdue watch item IDs reviewed.
• Goal check: Yes. The run found that adaptive agency needs both commitment and revision, and identified MAGELLAN as a genuine but narrowly scoped goal-selection mechanism. It also clarified why neither finding currently warrants implementation.
• Recommendation check: No material change survived. The no-action outcome is concrete, non-circular, bounded, approval-aware and better than adding an unvalidated self-scoring layer. Any future intervention would require an observed failure, external outcome measure, success criteria, rollback and separate approval.
• Tool-call failures: Capability gap — ScienceDirect returned a browser-restriction page rather than the review. Recovery: inspected the PubMed-indexed abstract, labelled the evidential limitation, and did not claim access to the full article.
• State updates: Added five source-index entries; reinforced the active 3.1 reflection that goal-revision and commitment-bias searches produce better signal than generic prioritisation queries; advanced rotation state to 3.2; wrote this report. No watchlist, backlog, experiment, disagreement or decision record changed.
• Stop reason: Five searches and five depth-inspected sources were sufficient to answer the focus. The next useful step would require local failure evidence or a new evaluation mechanism; inventing one would exceed the evidence and potentially touch protected process or system boundaries.