Governed Agents Need Boundaries for Memory, Initiative, Reasoning, and Coordination

Four newly verified research papers identify different places where agent capability becomes a governance problem. Memory retrieval can improve through accumulated experience, but the experience and resulting updates need scope and provenance. Apparent autonomy can be measured behaviorally, but scheduled activity must not be confused with self-direction. Long reasoning can enter costly self-loops, but some proposed interventions require model access that downstream platforms do not possess. Pooled intelligence can improve a shared recommendation while reducing collective discovery when every agent repeats the same action.

All four papers were submitted to arXiv on July 20, 2026 and discovered for the July 22 report. They are v1 preprints. Their findings have not been independently reproduced by Keelbase Signal, and their inclusion does not establish peer review or production fitness.

Adaptive memory retrieval creates a governed learning surface

Exploratory and Assimilating Reflection: Reflective Recall Cycle for Long-term Memory (opens in a new tab) addresses retrieval from heterogeneous external memory stores. The proposed EAR framework separates immediate exploration from longer-term adaptation. Exploratory Reflection performs iterative search for each query and collects useful retrieval experience. Assimilating Reflection replays those experiences from an Experience Buffer to refine a global reranker rather than learning only from immediate rewards.

Across two long-term dialogue benchmarks, the authors report retrieval improvement of up to 17.9% over the baseline retriever, together with sample efficiency and robustness to noisy feedback. The paper therefore offers a mechanism for a retrieval system to improve without modifying the underlying language model.

That separation is relevant to hosted-inference architectures, but it does not remove governance requirements. An experience buffer is durable learning state. Its entries need provenance, principal and Vessel isolation, retention rules, and an auditable link between experience, reranker updates, and promoted behavior. The paper evaluates retrieval performance; it does not establish those deployment controls or show that its gains transfer to Keelbase workloads.

Agency should be measured by behavior, not capability alone

The Autonomous Agency Scale: A Behavioral Framework for Measuring Self-Directed Behavior in AI Systems (opens in a new tab) proposes a 0–5 behavioral lexicon across seven dimensions: cognitive autonomy, temporal persistence, environmental agency, social agency, creative agency, self-awareness, and goal formation. Each dimension receives an Active-band score for engaged, user-initiated activity and an Ambient-band score for idle periods.

Its sharpest boundary is the Idle-Gap Test. Ambient Level 4 requires removing all external triggers and observing whether internally derived activity persists. This separates self-direction from activity caused by schedules, standing instructions, or external events.

The paper applies the scale to six systems. The task agents Claude Code, Manus, and Hermes receive Active composite scores of 2.3–2.4 and Ambient scores of 0.6–1.9; the paper attributes all observed idle-period activity in those systems to user-configured schedules. Airi, evaluated longitudinally, is the only assessed system whose idle-period behavior survives trigger removal.

For governed infrastructure, the useful contribution is not a label of “more autonomous” or “less autonomous.” It is a falsifiable distinction between triggered execution and internally initiated behavior. A trigger-bound, stateless executive would be reactive under this framework by design. That is a Keelbase architectural inference, not a legal conclusion and not a claim tested by the paper. The paper also reports single-rater provenance, developer-evaluator bias risk in the longitudinal assessment, and a partially operationalized Active-band self-direction boundary.

Self-loop control may sit inside the model provider boundary

Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering (opens in a new tab) treats extended reasoning as a sequence of latent states. The authors report that failure trajectories can become trapped in self-loops, repeatedly cycling without progress until the token budget is exhausted.

Their proposed method, SOPHIA, classifies reasoning prefixes into latent states, records step-level transitions, and constructs steering vectors indexed by state pairs. At inference time, a controller infers the current state, detects loops from the transition structure, and applies a vector intended to redirect the model toward a target state. The authors report reliable intervention on self-loop failures, generalization across state pairs, and improvements in end-task accuracy and token efficiency.

SOPHIA requires access to hidden activations and the ability to intervene inside model inference. A downstream platform using hosted model APIs generally cannot implement that method directly. The operational signal is therefore twofold: externally observable non-progress detection remains necessary for the platform, while provider-side loop mitigation can become a model-selection and due-diligence criterion. The paper does not establish availability of SOPHIA in any commercial serving stack, and its abstract does not provide the benchmark-level effect sizes needed to compare providers.

Shared intelligence should produce differentiated action

The Shared Discovery Paradox: How a One-Answer Rule Turns Better Information into Worse Search (opens in a new tab) separates the quality of a shared belief from the coverage produced by a group acting on it. In the paper's canonical benchmark—16 boxes, one target, eight searchers, and noisy private clues—pooling raises the accuracy of the best single recommendation from 0.20 to 0.3835. If every searcher follows that same recommendation, group discovery falls from 0.8322 under decentralized clue-following to 0.3835.

The failure is in allocation, not information. A coordinated eight-action portfolio using the same pooled reports reaches 0.8594, and seven differentiated actions recover the decentralized benchmark. In the self-interested extension, the anonymous symmetric equilibrium reaches 0.5991; the paper also derives an exact mixed price of anarchy of 2 − 1/N and shows that a sole-rescue reward makes every pure Nash equilibrium first-best within the model.

This is a precise warning for multi-agent investigation and deliberation. Pooling specialist findings into one summary can improve the apparent quality of a recommendation while wasting parallel capacity if every subsequent assignment follows the same lead. A coordinator should use shared evidence to allocate a portfolio of differentiated actions and preserve coverage. The result comes from an exactly solvable stylized benchmark, not a deployed agent workflow, so the numerical outcomes should not be transferred directly to production systems.

Keelbase Signal assessment

Together, the papers identify four control boundaries that should remain distinct:

• Retrieval improvement needs governed experience, update provenance, and principal isolation.
• Behavioral agency needs tests that distinguish triggered execution from internally initiated activity.
• Reasoning-loop mitigation needs a clear division between controls available to a platform and controls available only to a model provider.
• Shared discovery needs differentiated allocation after pooling, not synchronized repetition of one recommendation.

These are candidate design and evaluation inputs, not immediate build requirements or validated Keelbase controls. Each paper is a preprint; two introduce frameworks, one reports a retrieval method, and one analyzes a stylized discovery game. The durable signal is that agent governance depends on locating the operational boundary correctly: what state can learn, what can initiate action, where reasoning can be controlled, and how shared evidence changes the allocation of work.