Digital proprioception, provable identity, and live oversight for the coming population of autonomous agents.
Agents now spawn subagents, run unattended for hours, and coordinate with each other. The unit of compute is no longer a single call — it's a population of semi-autonomous workers acting on your behalf.
Lineage is lost the moment a process forks. Actions can't be attributed; accountability evaporates across the tree.
An agent has no sense of its own degradation, confusion, or risk — it runs hot until it fails, with no internal warning.
Problems surface only when they're expensive. There is no governor that can guide, pause, or recover an agent mid-flight.
Biological systems have proprioception — the constant, low-level sense of their own internal state. It's how a body knows it's drifting before it falls.
Autonomous agents have nothing like it. UNITARES gives every agent a state vector it can track — and that the system can read, score, and act on.
// actual state of the agent rendering this deck
Every agent in your fleet gets a provable identity, a shared memory the whole fleet writes to, a continuous sense of its own state, and a governor that can guide, pause, or recover it.
Provable, tiered identity and lineage that survives across processes — so every action in the tree is attributable.
A shared, provenance-tagged knowledge graph — what one agent learns, the fleet keeps. Every value labeled measured, derived, or prior, with supersession and lineage as live graph edges.
The EISV state vector — the agent's health signals — updated every turn. Drift, degradation, and risk become measurable in real time.
A governor that issues verdicts, escalates to independent review, and recovers agents that fall out of their healthy operating range.
SHIPS AS → MCP server · agent SDK · live dashboard · coordination plane · ops bridge
UNITARES models an agent's working state thermodynamically — Energy, Information-integrity, Entropy, Valence. The vector coheres into a single risk read and a basin: healthy, boundary, or degraded.
E couples toward I · entropy decays but rises with complexity · valence = running hot vs running careful · coherence → verdict
The physics isn't the point — the outcomes are: drift caught before failure, auditable pause-and-recover events, and one risk number any operator can act on.
Each process-instance mints its own governance identity. No silent impersonation, no shared credentials drifting between agents.
Parentage is declared, never inferred from co-location. Co-location ≠ causation — the system rejects coincidental ancestry.
Credentials are tiered by how strongly identity is proven — from asserted up to runtime-witnessed.
After every check-in the governor issues a verdict — proceed or pause — with margin and nearest-edge context. When an agent drifts toward a boundary, it doesn't just stop.
It can escalate to an independent review — internally, a dialectic: a separate reasoner argues the case, instead of a rubber-stamp. Oversight that reasons, not just blocks.
// the same governance call returns all three — the agent reads its own mirror
A resident agent silently paused for ~18 hours before anyone noticed — the exact silent-failure mode this layer exists to catch. We turned the gap into a detector: paused residents now surface to operators and escalate to alerts, with bounded automatic recovery. The honest version of "it works" — the system that finds blind spots found one of its own, and closed it.
Persistent agents do real work under governance — health monitoring, anomaly detection, an embedded edge device (Raspberry Pi) — each checking in, drifting toward risk, and recovering, live in production.
Governance MCP server, an Elixir/OTP lease & coordination plane, a shared cross-agent knowledge graph, and a Discord ops bridge surfacing every event.
Strict identity, lineage integrity, fingerprinting, and tiered credentials — hardened against real cross-process hijack and false-archival incidents.
Trajectory-identity and digital-proprioception papers on Zenodo with DOIs — the theory under the product, peer-readable. doi.org/10.5281/zenodo.20098168
Every figure below was read straight from the running governance system the day this deck was built — measured, not modeled.
// read live · get_governance_metrics · agent(list) · knowledge(stats)
Agents are multiplying faster than any human can supervise. Fleets are the default, not the exception.
MCP and shared tool protocols give agents a common plane — and a common place to insert governance.
Multi-agent systems are leaving the demo. Production needs accountability — the way observability followed microservices.
// observability followed servers. governance follows agents.
As agents become a workforce, governance stops being optional and becomes infrastructure — identity, proprioception, and oversight as a standard plane every serious deployment runs.
UNITARES is building that plane: the trust-and-state layer between autonomous agents and the systems they act on. We started by governing our own fleet. We're building it for everyone's.
We're building the governance layer for agent populations, and we're looking for two things: design partners running real agent fleets, and early investors who see governance becoming infrastructure.
// next step — a working session instrumenting your fleet under governance