A smaller controller helps AI agents reuse advice with less overhead

An agent using tools must decide what to do and respond to what happened in earlier steps. Reusing experience—such as a reminder about a required API argument—can help. But a recovery system may spend time generating a diagnosis before retrieving advice. And a lesson about the right tool or task can still be unnecessary or conflicting in the current state. The distinction is between topical relevance and present suitability.

Synapse-Control separates advice management from task execution. The main model, called the actor, still plans and takes actions. A lightweight control plane looks for situations where intervention may be needed, evaluates candidate advice, and governs whether to pass a hint to the actor.

How the control system works

Rules first use execution evidence to check for a possible need for advice—for example, a rejected API call, repeated unchanged results, or a visible gap in required information. If this deterministic check is triggered, rules shortlist up to five suggestions. A 1.7-billion-parameter classifier scores each suggestion against the current state as helpful, irrelevant, or harmful. The classifier was fine-tuned using labels from a larger model; it assesses candidate advice rather than generating a verbal diagnosis. A suggestion must pass a suitability threshold, and further delivery checks apply.

The reusable instructions are deliberately narrow. “Micro-contracts” capture local interface constraints, such as required arguments or valid options, rather than prescribing a complete sequence of actions. The actor remains responsible for choosing how to accomplish the task.

A Prompt Value Table records what advice has already been delivered and feedback about what followed. It uses evidence changes, cooldowns, and a token budget to limit repeated hints. Advice is offered as a temporary prompt for the next action, while the resulting actions and observations remain in the task history. The existing recovery host continues to handle verification and updates to reusable lessons.

What the tests found

The authors tested Synapse-Control on AppWorld, a benchmark for stateful API programming, and WebShop, a simulated shopping environment. They compared it with an unassisted agent (Base) and Sentry, a system that generates diagnoses before selecting advice. Each method had 1,308 registered attempts: 1,008 in AppWorld and 300 in WebShop. The primary comparison used the same actor and task workload. Some AppWorld attempts were not scored, but completion and cost totals include all registered attempts.

Synapse-Control completed 270 attempts, compared with 256 for Sentry and 208 for Base—29.8% more completions than Base on this workload. That count should not be confused with mean benchmark score: Synapse-Control’s mean scores were slightly below Sentry’s in both benchmarks. In a separate comparison matching the knowledge available to the systems, Synapse-Control’s success rate was higher in AppWorld but lower in WebShop than the generative control’s.

The clearest efficiency result concerns auxiliary control time: compared with Sentry, it fell by 50.2% in AppWorld and 51.0% in WebShop. Synapse-Control replaced 4,987 generative diagnosis calls. Total task time was also lower than Sentry’s in the primary comparison, but the reduction varied: 5.0% in AppWorld and 21.4% in WebShop.

A component test points to the role of delivery history. When the Prompt Value Table’s history checks were removed, repeated guidance deliveries rose from 23 to 65 in AppWorld and from 2 to 6 in WebShop. The classifier also achieved high precision on held-out micro-contract candidates against teacher-provided usefulness labels. Those test results concern agreement with labels, not proof that each suggestion improves task outcomes. In the Contract-candidate tests, 50 of 76 AppWorld candidates were accepted, and one of those 50 accepted candidates was labeled harmful; in WebShop, 108 of 150 candidates were accepted, with none of those accepted candidates labeled harmful. Other mixed-content tests included harmful acceptances.

What remains uncertain

These are results from two benchmark environments, not evidence that the approach will work equally well across everyday software or other domains. The experiments use a particular actor model and domain-specific interface resources. Porting the system requires adapting observations, providing local interface facts, and calibrating the controller. The authors estimate that organizing one domain’s templates takes a few person-hours, but the evaluation does not establish the effort or reliability of onboarding arbitrary real-world systems. Offline labeling and training also add costs not captured by online control-time measurements.

The paper is an arXiv preprint and may not have undergone peer review. Independent replication and testing in other settings would help establish how broadly the findings apply. The supported takeaway is narrower: in these AppWorld and WebShop workloads, the system reduced measured auxiliary control time relative to Sentry, while its completion counts and benchmark scores did not uniformly outperform that comparison.