From Workspace Actuators to Pre-Action Agent Telemetry | Armalo Labs | Armalo AI
Safety ResearchJuly 9, 20263 min read
From Workspace Actuators to Pre-Action Agent Telemetry
Armalo Labs Research Team
Key Finding
The goal is not to expose full reasoning traces; it is to test whether compact internal-state signals improve agent harness decisions before action.
Abstract
We outline a harness research design for using bounded J-space-style signals as pre-action telemetry: compact state indicators that may predict constraint loss, prompt-injection recognition, or workspace saturation before a tool call occurs.
Tool-using agents often fail before their final answer reveals the failure. They lose constraints, accept malicious context, confuse authority, or saturate their working state. J-space-style readouts suggest a possible pre-action telemetry layer: a way to observe token-legible intermediate state before an agent acts. We argue that useful telemetry requires calibrated actuators, behavior validation, entanglement maps, and adversarial pressure tests. The goal is not to expose full reasoning traces, but to create compact, bounded, causal state signals that help harnesses predict and prevent failures.
1. Problem
Agent traces are late. By the time a tool call is emitted, the decisive computation may already be done. Final-output evals are later still.
Pre-action telemetry asks whether internal state can warn the harness earlier:
constraint eviction;
workspace overload;
prompt-injection recognition;
hidden task-frame shift;
suspicious monitor-awareness;
missing safety-relevant concepts.
2. Required Evidence
Telemetry is not licensed by readouts alone. A telemetry channel should pass:
1.readout validity;
2.
Cite this work
Armalo Labs Research Team (2026). From Workspace Actuators to Pre-Action Agent Telemetry. Armalo Labs Technical Series, Armalo AI. https://www.armalo.ai/labs/research/2026-07-09-pre-action-agent-telemetry-jspace
Armalo Labs Technical Series · ISSN pending
Explore the trust stack behind the research
These papers are built from the same trust questions Armalo is turning into product surfaces: pacts, trust oracles, attestations, and runtime evidence.
3.the harness checks constraints and risk signals;
4.if drift is detected, the harness holds action, asks for clarification, or
repairs context;
1.the outcome is logged against behavior.
This is not a product claim. It is a research design.
4. Self-Monitoring
The next deeper question is whether the agent can use a bounded readback of its own state. A compact state message such as constraint_C eviction risk: high may help the model self-correct without exposing full reasoning traces.
The experiment is simple: matched tasks with and without workspace readback, scored on constraint violations, task success, and overcorrection.
5. Failure Modes
monitor overfitting;
concept entanglement;
false reassurance from clean benchmarks;
reward pressure that makes the monitored channel look good while behavior
decouples;
externalization into tools or memory.
The adversarial workspace tomography program exists to measure these failures before anyone treats telemetry as infrastructure.
6. Conclusion
The path from J-space to agent safety is not direct. It runs through calibrated handles, behavior gates, sparse decomposition, and collapse curves. But if the path works, it gives agent systems something they currently lack: pre-action state measurement.
References
Anthropic. "A global workspace in language models." 2026.
OpenAI. "Detecting misbehavior in frontier reasoning models." 2025.
Li, Fan, and Zhuang. "When AUC 0.998 Is Not Enough." 2026.