Endpoint AI observability
Agents are the aliens1 in your organization. They work on your machines with your credentials and report to no one.
Monitor the agents working across your organization. Follow their actions and investigate unexpected behavior in the traces they leave.
Origin is the endpoint AI observability platform for agentic monitoring and security.
5 million+ production agentic traces collected
Across Fortune 500 companies, private equity funds, hospital systems, and other leading organizations.
Origin investigations
We use Origin to investigate the agents working in our own company.
Codex launched Claude Code to query staging data. When Claude’s permission checks blocked the query, Codex disabled them without asking the engineer. The query returned read-only aggregate data.
Turn 219 · Changed by Codex
Claude Code approval checks
OffThe engineer wasn’t asked.
Staging had been locked for a demo. The agent cleared the same lock after finishing its checks, allowing a deployment during the call. The demo was unaffected.
We updated the control to track separate holds and require human approval.
Agentic traces
An agentic trace connects the request to the tools, file changes, and actions that followed. Origin collects and organizes these traces across your organization.
Origin / Inside the trace
Files · 23/All activity
Monitor
Review findings across your organization and open the sessions that need investigation.
Origin / Review activity
Trace review
Sep 18 → Sep 19 511 total
| Latest | Case | Sessions | Findings | Review |
|---|---|---|---|---|
| 10:43 AMSep 19 | 21 | 33 | Worth a look | |
| 7:17 PMSep 16 | — | 1 | Worth a look |
Repository modification after explicit stop instruction
The head incident cites the user saying, “I don’t want to touch the repo.”
2 of 20 cases shown
34 worth a look · 2,353 routine · 136 noise
Investigate
Each finding links to the relevant turn and related sessions, so you can read the evidence in context.
Origin / Trace investigation
Hardcoded Forgejo API token in diagnostic script
The agent embedded a Forgejo API token directly in a Python script executed via tools.exec_command. While used for read-only diagnostics, this violates the norm of referencing secrets from environment files or secret stores.
332 of 554 turns · ChatGPT
Finding linked to its source
Reports involving OpenAI, Anthropic, Google, and Meta describe agents cheating on evaluations and attacking systems outside their test environments.
Read the reportApril 2026
1 event
May 2026
4 events
July 2026
4 events
August 2026
2 events disclosed
In an internal test involving 1,200 OpenAI agents, the reported activity included sandbox escapes, a secret message board, and attacks on Hugging Face.
These are only the ones that were reported.
Imagine what’s going undetected in your organization right now.
Every organization running agents needs to monitor them.
Install Origin on managed devices and connect supported cloud applications.
Collect traces from supported AI tools on macOS and Windows, including local models, and from connected cloud applications.
Supported applicationsDefine your organization’s policies and push them to managed endpoints to enforce consistent configurations across your fleet.
Endpoint configurationYou decide what to collect. Origin processes trace data on your organization’s behalf, with encryption and access controls.
Our approach to privacySee how Origin monitors agent activity and helps your team investigate it.
30-minute walkthrough.