Claude Code
Track shell/tool cycles, repeated edits, long retry chains, context growth, and cost exposure in coding sessions.
Agent coverage
One structural meter across the agents already doing the work.
CAUM watches runs from coding agents and automation agents for loops, repeated tool cycles, stalled progress, token exposure, and cost exposure. It observes structure only. It does not read prompts, source code, or private content.
Agent families CAUM can observe.
Different agents emit different logs. CAUM normalizes the shape into structural events, then reports the health of the run without scoring the content itself.
OpenAI Codex
Observe coding-agent runs from event logs, terminal activity, patch attempts, validation loops, and token or cost metadata.
Cursor / Windsurf
Convert editor-agent activity into structural evidence when session exports, tool traces, or automation logs are available.
OpenHands
Measure action/observation rhythm, command repetition, environment retries, and paths where compute stops becoming movement.
SWE-agent
Summarize issue-solving traces into structural health tiers and hard-alert evidence for exact loops and dead-work pressure.
LangChain, AutoGen, custom
Send neutral JSON events from your orchestration layer and keep prompts, files, customer data, and model output private.
Real validation agents are now running CAUM.
CAUM is being exercised by real non-destructive processes: production website smoke checks, backend tests, trace ingestion, privacy-boundary probes, and a controlled retry loop. The loop control is expected to alert; healthy agents are comparison runs.
Latest production validation
April 30, 2026 run against the live Railway API. Honest read: the controlled retry loop reached T5 and produced a live alert; non-control agents completed and reached review tiers, which keeps T4 labeled as review-only, not public waste.
5
real process agents
T5
controlled retry loop
0
private canary leaks
4
comparison workflows
Private content stays outside the meter.
CAUM needs the structure of work, not the meaning of the work. Teams can hash identities, omit prompts, and send only operational signals that describe what happened during the run.
{
"run_id": "run_042",
"agent": "claude_code",
"event_type": "tool_call",
"tool_name": "shell",
"status": "retry",
"tokens": 1840,
"cost_usd": 0.021,
"timestamp": "2026-04-30T14:25:00Z"
}
Two ways to start.
Use the PDF Receipt when you already have logs. Use Live Meter when you want observability during the run.
1. Upload
Drop messy JSON, JSONL, or exported trace data into the PDF Receipt flow and receive a structural audit after the run.
2. Stream
Send live event records from your agent runtime and watch structural health, hard alerts, tokens, and cost exposure.
3. Act
Use CAUM evidence to review wasted compute, tune retries, compare agent workflows, and decide what deserves deeper inspection.
CAUM is the meter between agent activity and actual progress.
It gives operators a structural receipt for where compute went, without pretending to know whether the agent was semantically right.