Blog · arXiv Analysis · Published: June 25, 2026 · Modified: July 10, 2026 · Last reviewed: July 10, 2026

The Agent Action Becomes the State Signal

Andres Enriquez Fernandez and John J. Bird's June 2026 paper asks a concrete coordination question: when an autonomous agent cannot reliably talk, can its control policy make its actions easier to read as evidence of its state?

An action-state signal is the information about an agent's hidden or partially hidden state that can be inferred from the actions it takes. The governance problem is deciding when to train for that legibility, who is allowed to read it, and what record proves that legibility did not become a new leakage channel.

The Paper

The paper is Training Observable Control Policies to Expose Agent State Through Actions, arXiv:2606.27609 [cs.LG]. arXiv lists it as submitted on June 25, 2026, with cross-listing in Systems and Control, a related DOI of 10.2514/1.I011654, and a journal reference in the Journal of Aerospace Information Systems. The authors are Andres Enriquez Fernandez and John J. Bird.

The site already has pages on agent observability through entropy traces, context dashboards as agent proprioception, and process traces as accountability maps. This paper adds a different object: not a log after the fact, but the control policy itself as a source of observable state information.

Current Context

As of this July 10, 2026 review, the arXiv record identified the paper as version 1 and listed the journal reference in the Journal of Aerospace Information Systems. The result should be read as a simulated aerospace-control study, not as a general certification that autonomous agents are human-interpretable, secure, or safe in every communication-limited environment.

The governance context is converging on the same practical question from a different direction. NIST's AI Agent Standards Initiative frames agents capable of autonomous actions as a standards problem involving trusted, interoperable, secure operation. The April 2026 allied guidance Careful adoption of agentic AI services warns that agentic systems can be difficult to understand, monitor, audit, and reproduce, and recommends monitoring all agent operations, not only inputs and outputs. EU AI Act Article 12 requires high-risk AI systems to support event logging appropriate to purpose, and Article 14 requires human oversight measures that let assigned humans understand capacities and limits, monitor operation, interpret outputs, override or reverse outputs, and interrupt the system where appropriate.

Those sources do not validate Fernandez and Bird's controller. They show why the paper matters: in some systems, logs and dashboards are not enough. The action sequence itself may be part of the observability design, and that creates both safety value and information-leakage risk.

Action Channel

The paper starts from a constraint familiar to anyone who has worked with distributed machines: direct communication can be delayed, dropped, unavailable, operationally risky, or absent because systems do not share hardware or language. In those cases, an observer may still see what an agent does. A vehicle brakes, an aircraft changes course, a robot alters its path. The action is not a message, but it can carry information.

Fernandez and Bird formalize that intuition. They describe an agent as a control policy mapping state to a distribution over actions. If an observer knows the policy and system dynamics, observed actions can be used as measurements for estimating hidden state. The paper uses an Unscented Kalman filter for that estimation problem, with the agent's sampled pseudocontrol actions as observations.

That framing changes the governance question. Monitoring is often treated as a separate interface: add telemetry, add logs, add a dashboard. Here, the policy is trained so that ordinary actions become more informative. The legibility is not bolted onto the side. It is part of the behavior that the agent learns.

The phrase "state signal" should be kept narrow. It does not mean the action reveals intent, mental content, or a reliable explanation for every decision. It means the observer can use the action history, together with a policy model and dynamics assumptions, to improve estimation of selected state variables. If the policy, dynamics, sensors, or observer assumptions drift, the signal can become stale or misleading.

Training Loop

The authors compare a task-only policy against a policy trained with an embedded estimator reward. The task-only policy is rewarded for the aircraft tracking task. The embedded-estimator policy is rewarded both for the task and for making the estimator's state reconstruction better. The paper is careful about the tradeoff: a reward bonus for estimator performance could reduce task performance, so task reward must be measured separately.

This is where the work becomes relevant beyond aerospace. A warehouse robot, field drone, software agent, or vehicle fleet can appear competent while being hard to monitor. A policy that maximizes task success alone may converge on behavior that hides state from observers, even without any deceptive intent. An observability term says that success includes leaving enough behavioral evidence for coordination and review.

That design move is powerful and dangerous for the same reason. Training for observability changes behavior. The agent may take a slightly different route, steer differently, reveal a pattern earlier, or make its state easier for a teammate to estimate. In a hostile setting, the same pattern may help an adversary infer location, target, operational mode, or future maneuver. Observability has an audience, and that audience must be governed.

Aircraft Test

The experiment uses a fixed-wing uncrewed aircraft system assigned to remain near a ground target by controlling bank angle. A second estimator tries to infer the aircraft's position and velocity relative to the target from the steering command. The paper evaluates the task-only and embedded-estimator policies over an ensemble of 14,500 episodes.

The reported state-estimation gains are substantial. At 200 seconds, the mean position-error norm is 21.77 meters for the task-only case and 12.15 meters for the embedded-estimator case, a 44.2 percent decrease. The mean velocity-error norm falls from 4.33 meters per second to 2.94 meters per second, a 32.1 percent decrease. The task penalty is small in the reported comparison: at the 80th percentile, task rewards are 80.1 for the task-only case and 79.1 for the embedded case, a 1.2 percent difference.

The observability analysis is also important. The authors report that neither policy is fully observable at any single state point. The difference appears over a trajectory. Using a stripped observability matrix, they find that after about 50 seconds, roughly two cycles, the embedded case shows a stronger sequence-level observability signal. The result is not "the state is always readable." It is narrower: the learned action sequence can make estimation easier over time.

Failure Modes

Legibility overclaim. A system reports that a policy is observable without specifying which state variables are estimable, over what time window, under which dynamics, by which observer, and at what error threshold.

Estimator privilege. The action sequence is readable to an Unscented Kalman filter with policy and dynamics knowledge, but not to the human operator who must coordinate with the agent during workload, latency, stress, or partial observation.

Adversarial readability. A teammate can estimate state from actions, but so can an adversary watching the same maneuvers, tool calls, file edits, network requests, or route choices.

Task-legibility tradeoff drift. A small penalty in simulation becomes an unacceptable mission, safety, energy, privacy, or latency cost in a different environment.

Policy mismatch. The observer estimates state using an old policy model after the agent has been retrained, fine-tuned, patched, prompted differently, or given a new runtime wrapper.

Silent opacity. Operators believe action legibility is available, but sensor loss, missing actions, stochastic sampling, changed dynamics, or environmental disturbance makes the estimate unreliable without warning.

Governance Standard

A deployed observable-policy system should treat action legibility as an explicit safety property, not an informal hope. The system owner should name the estimable state variables, intended observers, permitted uses, error bounds, update cadence, stale-policy alarms, adversary model, and conditions under which the action-state signal is not reliable enough for coordination or oversight.

The governance record should also separate three cases: machine-estimable, where a formal estimator can reconstruct selected state variables; operator-legible, where trained humans can interpret behavior well enough for coordination; and audit-legible, where a later reviewer can reconstruct the sequence from records. A system can satisfy one without satisfying the others.

In agent software, the same logic applies outside aircraft. A coding agent's file edits, a browser agent's clicks, a payment agent's mandate requests, or a warehouse robot's route can all signal internal state or task interpretation. Those signals belong with agent logs, action certificates, and runtime governance: the action is evidence only when the policy, context, observer, and limits are recorded.

State Receipt

An observable-policy system should leave a state-estimation receipt. At minimum, the record should include the policy version, state variables intended to be estimable, action variables exposed to the observer, estimator type, assumed dynamics, reward terms, observability metric, training environment, test distribution, task-performance penalty, divergence rate, and reviewer threshold for unacceptable opacity.

That receipt matters because action legibility is a double-edged design feature. A signal useful to teammates may also be useful to adversaries, competitors, supervisors, or insurers. A governance process has to ask who is authorized to read the state signal, whether the signal leaks sensitive operational intent, and whether the agent can continue safely when the observer's model of the policy is stale.

For human-machine teams, the receipt should be even stricter. A Kalman filter can process a policy in a way a human cannot. If the goal is human understanding, the system needs evidence that actual operators can interpret the behavior under workload, stress, latency, and partial observation. Otherwise "observable" only means observable to a mathematical estimator with privileged model knowledge.

For consequential deployments, the receipt should include a stop condition. If position or velocity error, covariance, divergence rate, missing-action rate, policy-version mismatch, or operator disagreement crosses a threshold, the system should degrade, escalate, communicate directly, or hand off rather than continue as if the action channel is still trustworthy.

Claim Boundary

The paper does not claim to solve human trust, general autonomy, or broad agent governance. It studies a specific estimation problem under a simulated aircraft tracking task. It also names a human-machine limitation: human understanding of system state is unlikely to follow the same dynamics as the Unscented Kalman filter used in the experiment.

That boundary is the useful part. The paper gives governance a precise design question: before deploying an agent into a communication-limited environment, can its actions be read well enough to support coordination, auditing, and safe handoff without destroying task performance? If the answer is no, the agent is not merely quiet. It is operationally opaque.

Source Discipline

Use Fernandez and Bird's paper for the action-as-observation frame, the embedded-estimator reward design, the UKF-based estimation setup, the fixed-wing aircraft simulation, the reported 14,500-episode evaluation, and the authors' stated limits. Do not generalize it into a proof that action legibility solves human trust, multi-agent safety, military coordination, robotics assurance, or software-agent monitoring.

Use NIST, allied agentic-AI guidance, and EU AI Act Articles 12 and 14 as governance context. They support the broader need for monitoring, logging, oversight, interpretation, interruption, and risk management. They do not certify this controller, define a universal state-receipt standard, or prove that an estimator-readable policy is human-legible.

The clean claim is narrow: when communication is limited, an agent's action sequence can be intentionally shaped to improve selected state estimates. That can help coordination and audit, but only if the authorized observer, leakage risk, model staleness, task penalty, and human interpretability limits are part of the receipt.

Sources

Return to Blog