Census

The coherence census collects reads of real agents. Each row carries the trace the read consumed, the mapping from that trace to committed state, and the exact command that produced the read, so anyone can run the row again and reach the same finding. The census lives at github.com/RightRudderAI/coherence-census.

Full runs

These agents carry a complete run

DBOS Hacker News agent

In our reread of DBOS's own Hacker News research agent, vendored unchanged and run on gpt-4o-mini through OpenRouter across five topics at ten iterations each on 2 September under fathom-read 0.4.1, the agent as it shipped repeated 14 of its 50 searches, re-read 42 percent of the threads it fetched, and held 516 stories more than once. With the repair in front of the step where the agent proposes its next queries, the repeats fell to 0 of 50, the re-reads to 12 percent, and distinct threads covered rose 35 percent, at the same model and the same iteration count.

CrewAI book flow

We read CrewAI's book-writing flow, an outline crew and per-chapter crews on CrewAI 1.15.21, run on three models through OpenRouter at three runs each and captured under fathom-read 0.3.0 in September 2026. On gpt-4o-mini the flow read coherent across all three runs. On llama-3.3-70b-instruct it drew duplicate commits in bulk, 28 in one run and 21 in another, along with two stale references, and the span read matched the event-bus read finding for finding. On mistral-small-3.2-24b it read coherent but for two duplicate commits in a single run.

DeepAgents deep research

We read a DeepAgents deep-research run on deepagents 0.7.14 with langchain-fathom, an orchestrator delegating to a sub-agent through task calls with Tavily search, on llama-3.3-70b through OpenRouter, asked to compare five long-term memory systems. The run made 6 tool calls, held one source, and delivered its report with a single citation. The read returned one stale reference, since the report cited a source whose URL slug it had reordered from the article the committed state actually held.

Earlier reads

Earlier reads across frameworks

The census grew out of reads we ran on production frameworks, each showing where an agent's own record and its reported outcome part ways.

Agent Zero Memory
A provenanced-memory agent passed its own citation lock on none of 83 stale answers, and the committed-state read caught every one.
ContextPilot
With the agent managing its own context, 55 to 60 percent of questions returned a superseded value, and the models recovered the update in fewer than one run in ten.
LangGraph
The read reconstructed a coherent record across more than 170 checkpoints and caught the inconsistency when two agents wrote at once.
CrewAI event stream
Sequential crews held coherence, while a hierarchical crew reported every task complete having written no records.
Letta
The read recovered stale archival passages the agent's visible edits had missed.
OpenHands
A small-model run failed to make its edit yet reported success, and the read separated the actual outcome from the reported one.
Agent-E
Under window pressure, the agent silently duplicated orders while its dashboards stayed green, and the read caught every duplicate.
Kosmos
A summary world model carried a coherence-cost signature that stood apart from task difficulty.
Every row reads from the record a run already produced. The census shows the failure and the measurement. The repair that fixes a run stays in the hosted service.