Research

The evidence.

Case studies on third-party agents and frameworks, controlled experiments and benchmarks, and the program papers underneath them.

Newest first within each group.

Third-party agents and frameworks

The committed-state read on systems ERA did not build

Each study runs the read, unchanged, over what the framework or agent already emits: checkpoints, event streams, persisted memory, edit logs, tool returns.

Case study · September 2026New

Reading committed state when memory is provenanced

A Fathom coherence read run against the reading discipline of Agent Zero Memory, on LongMemEval.

Agent Zero Memory (Zhu and Wu, 2026) gives every stored item its origin, timestamp, and evidence pointer, and holds answers under a citation lock: a reply may cite only evidence the reader opened, or it abstains. We implemented that reading discipline as the paper states it and ran it on LongMemEval's knowledge-update questions across deepseek-chat and gpt-4o-mini. When the update fell outside the retrieval window, the reader cited the superseded item and answered with the superseded value, and every one of those answers passed the citation lock; across 83 stale answers on the two models, the lock rejected none. A timestamp check on cited items would catch most of them, but not the answers that cite both the old and the new item and commit the old value. The committed-state read fired on every stale answer and stayed silent on the coherent ones.

Case study · September 2026New

Reading committed state when an agent manages its own context

A Fathom coherence read run against a ContextPilot-style context-management step, on LongMemEval.

We ran the committed-state read against ContextPilot's own foldHistory operation (Pan et al., EMNLP 2026, Apache-2.0) on LongMemEval's knowledge-update questions, across deepseek-chat and gpt-4o-mini. Once we folded the update, roughly 55 to 60 percent of questions came back with the value the agent had already replaced, the models searched the folded index to recover the update on fewer than one in ten runs, and the read produced no verified false positives across more than a hundred runs. A second experiment shows a graded answer passing while a fact the question never touched was left stale, which is the case a single-probe score misses.

Case study · August 2026New

Reading a graph's committed state from its own checkpoints

A Fathom coherence read run against LangGraph.

LangGraph persists every value an agent commits as an ordered lineage of checkpoints. We reconstructed the committed record from that lineage alone, with no ground truth, across three coherent multi-agent runs and more than 170 checkpoints, where the read stayed silent. On a single parallel super-step where two agents wrote the shared record at once, the channel reducer merged both writes with no consistency check, and the read recovered the record left citing the old key.

Case study · August 2026New

Reading a crew's committed state from its own event stream

A Fathom coherence read run against CrewAI.

We gave a CrewAI crew a coupled record-keeping job across 5, 12, and 20 interdependent sub-records under sequential and hierarchical processes, on two models. We reconstructed committed state from the events the framework already emits and each tool call's success flag. The sequential crews held the record coherent and the read stayed silent; a hierarchical crew reported all twenty sub-tasks complete while it never wrote one record, and the read recovered the dropped record from the event stream alone.

Case study · August 2026New

Reading an agent's committed state from its own memory

A Fathom coherence read run against Letta.

A Letta agent's committed state is its memory: core blocks edited in context and archival passages written out of context. With a fact carried across eight core blocks, the agent renamed every copy and the read stayed silent. With six core blocks and one archival passage, the same model renamed the six it could see and left the archival copy on the old key. Reading only the persisted memory and the agent's own edit stream, the read recovered the stale passage.

Case study · August 2026New

Committed-state coherence in a coding agent

A Fathom read on OpenHands.

We asked OpenHands to rename an identifier across a five-file project and run the tests, under a capable model and a small model with a distraction step. The capable run committed the rename across all five files and read clean. The small-model run made no successful edit, ran the suite against the unchanged code, and reported the task complete with tests passing. The committed-state read, reconstructing the repository from the agent's own edits and success flags, was the one check that separated the reported success from the actual one.

Case study · August 2026New

Reading coherence failure in a live web agent

A capability-graded study of Agent-E under retained-window pressure.

Agent-E, the open-source web-navigation agent published by Emergence AI, ran an interaction-dense ordering task across two model tiers and a range of retained-window budgets, with three reads over its reconstructed committed state. At a fair window, a frontier model completed the order and every read stayed silent. With the window starved, the same model silently duplicated the order (8 lines and $106.20 against the correct 3 lines and $44.10) while every conventional success signal stayed green, and the duplicate-state read recovered all five duplicates from the action stream alone.

Case study · August 2026New

Coherence cost in a summary world model

Separating an AI scientist's synthesis-accuracy gap into task difficulty and coherence burden, on a controlled surrogate of the Kosmos architecture.

Kosmos, the AI scientist published by Edison Scientific, reports cross-trajectory synthesis well below single-source analysis. On a controlled surrogate of its summary-world-model architecture, with ground truth fixed in code, we separate that gap into task difficulty and coherence cost against a matched-difficulty control. The summarization deficit carries the coherence-cost signature and the control does not; the cost that remains at full legibility shrinks with model capability without closing; a single bad committed value cascades through dependent quantities; and a self-verification pass does not remove it.

Controlled experiments and benchmarks

Experiments ERA runs, with ground truth fixed in code

Case study · August 2026

Coherence risk in long-horizon coding agents

Reading the binding cause of a coding-agent failure and applying the matched repair, on a controlled coding task with a low-cost model and a frontier model.

Two coding-agent failures present the same symptom, a change that compiles and passes tests and fails when run, and call for different repairs. One is a contradiction of the agent's own earlier commitment; the other is a step that is simply too hard. A selector reads which cause binds from the action stream, routes re-grounding or escalation, and declines the mismatched repair. Executing the committed code grades every result, on a low-cost model and a frontier model.

Case study · July 2026

Coherence risk under coupling

A controlled Fathom experiment on Cloudflare Workers AI: when an agent's state loses coherence, and how a committed, auditable record restores it.

When an agent maintains coupled state, where each value it tracks depends on others, coherence becomes a question of how strongly those values feed one another and whether the record can keep up. On a controlled task running live on Cloudflare Workers AI, past a coupling threshold the agent's state loses coherence whether the agent holds it in the conversation or in an external store. Re-grounding the record to truth arrests the failure, and the audit can be sparse: re-grounding as rarely as one turn in ten holds the system below the point of collapse.

Benchmark · July 2026

The real cost of agent memory

An instrumented, cross-cloud benchmark of managed memory on Amazon Bedrock and Microsoft Azure.

The same benchmark run on both clouds where enterprises field agents. In the default configuration, the agent loses commitments it made earlier in the run and contradicts them. The premium managed-memory path holds those commitments well, and re-grounds the agent on each one through a model-generated summary whose token cost climbs with accumulated state: recovering one committed fact after a hundred commitments cost about 155 times a deterministic lookup on Bedrock AgentCore and 85 times on Azure AI Foundry, and neither path leaves a record an auditor can cite.

Measurement study · August 2026New

Reasoning is the default modality, and it breaks naive agent measurement

A Fathom measurement study across five frontier model vendors.

We probed eight models across five vendors on a trivial and a harder task at tight and ample output budgets, with reasoning left default, explicitly off, and explicitly high. On four of five vendors, the flagship reasoning endpoint reasons by default and rejects an explicit request to disable it; where hidden thinking shares the response budget it can spend the budget and return nothing, which an ordinary harness records as a wrong answer rather than a broken measurement. The probe ships with the study, so every figure is reproducible.

Whitepaper · August 2026New

The rise of stateless agents

Why the industry moved to memoryless swarms, what that buys, what it caps, and the measurable tipping point between sharding work and carrying state.

Agents that run long enough to be useful have to remember what they have done, and remembering is where they break. Much of the industry has moved to swarms of small workers that remember nothing, coordinated by a thin controller. Under the same reliability condition, a stateless swarm carries a reflexive burden near zero, which confirms what the architecture buys, and it exposes a quieter risk: under load, the swarm stops backing up most of its own conclusions while every surface dashboard reads healthy. Letting workers build on one another's results, which is what adding memory does, switches the burden back on, creates a sharp failure boundary, and sets off cascading collapse, and the effect reproduces on two live production model families. The tradeoff has a measurable setpoint between accumulating and sharding.

Whitepaper · July 2026New

Context efficiency: the second lever for coherence in long-horizon agents

A specific and measurable portion of what an agent must store is the price of modeling its own behavior, and the harness can reduce that portion directly.

The dominant response to incoherence has been to enlarge the context window, which treats a structural problem as a capacity problem. This brief presents the first controlled evidence for a second lever: spending the window on a compact snapshot of the agent's own committed state rather than its whole history, so the same coherence holds on a smaller and cheaper context. On a controlled coding task, measured in exact billing tokens on Claude Sonnet 5, the managed agent held the same coherence on a fixed budget of 16,680 tokens per trajectory where the raw-window agent needed 113,636. The paper states where the result stands and what a deployment claim on real code still requires.

Program papers

The theory and the program

Whitepaper · July 2026

Predicting coherence failure in long-horizon AI agents

From the measurement of reflexive burden to the prediction of agent failure. The Fathom program whitepaper, v21.

The program paper. Why long-horizon agents degrade in a way single-turn evaluation cannot see; the feasibility inequality that separates task difficulty from reflexive burden; the instrument suite that measures, attributes, forecasts, and repairs coherence loss against matched controls.

Theory · preprint · SSRN 2026

Records, reflexive modeling, and the conditions for stable physical histories

A feasibility inequality and three minimal models.

The information-theoretic science behind Embedded Risk Analytics. 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, written C ≥ Hext + Bρ. The reflexive term is different in kind from ordinary difficulty: it grows as a system approaches its stability limit, where difficulty stays bounded.