About
Embedded Risk Analytics turns a theory of self-tracking stability into deployable middleware for AI agents.
ERA is an independent, founder-run venture. The Fathom program is its product: a measurement-and-repair harness that runs beside a live agent, reads where the agent loses coherence, attributes the cause, and repairs the reflexive part in line.
What we do
On a long task, an agent loses track of its own prior actions and begins to contradict itself. Leading platforms hold an agent's committed state in managed memory, and hold it well, through a probabilistic model summary at a token cost that grows with the state and without a record an auditor can cite. What no tool provides is the reading underneath: which cause is binding, and a deterministic account of the state the next action depends on. That reading is what ERA supplies.
Measure
Reads coherence loss as it happens, from the agent's own action and tool-return stream, with no model internals.
Attribute
Attributes each loss to reflexive burden (the agent's own carried state) or to ordinary task difficulty, against a matched control that runs the same task with known arithmetic in place of the model.
Repair
Applies the matched repair only where the measurement predicts it pays: re-grounding on reconstructed committed state where carried state binds, escalation or a verified tool where difficulty binds, and neither where neither helps.
The instruments are deterministic code over the action stream and add effectively no cost or latency. The matched-controlled dataset they produce, 15,995 controlled trajectories and about 714,000 labeled per-turn decisions across fourteen task families as of August 2026, is the venture's durable asset.
The theory
A system stays coherent only while its capacity covers the ordinary difficulty of its task and the burden of tracking its own effect on that task. We call the second term reflexive burden, and we write the condition as C ≥ Hext + Bρ. The theory shows the reflexive term is different in kind from ordinary difficulty: it grows as a system approaches its stability limit, where difficulty stays bounded, and that asymmetry is what difficulty-based reliability measures miss. The same accounting describes how AI agents come apart on long, multi-step tasks.
The preprint on SSRN sets out the theory, and the Fathom whitepaper sets out the program that operationalizes it.
Deployed where agents run
The harness runs live on Amazon Bedrock AgentCore and Azure AI Foundry, with the committed-state read available as a callable tool on each. On Cloudflare Workers AI, it runs a Durable-Object committed store. The committed-state read also runs offline over LangGraph state history, OpenInference traces from Arize Phoenix, CrewAI event streams, Letta memory, and the action logs of tool-calling and coding agents.
Integrations and the committed-state layer are open source at github.com/ERA-Fathom. Attribution and scoring run only in ERA's hosted service.
Intellectual property and stage
A USPTO provisional patent application is on file (64/118,017, filed July 2026) covering the measurement, attribution, anticipation, and repair of coherence failure in long-horizon agents through a shared, deterministically reconstructed committed state. The underlying theory appears as a preprint on SSRN.
ERA is opening design-partner and licensing conversations. The first step with any team is the same: an offline committed-state read on a sample of their own agent's traces, with a written readout. How to start →