Everyone already has an agent.
Why can't they work together?

An open-source runtime for long-running agentic systems, built on the world's first data-centric operating system. Bring your harness. Let it take an oathe.

RUNS ALL OF THESE. TOGETHER.
Claude Code Codex · ChatGPT Cursor OpenClaw Hermes Grok Bot GitHub Copilot Microsoft Scout
$ npm install -g @oathe/oathe
$ oathe init
shez@mac app % oathe claude
Claude Code v2.1
Fable 5 · Claude Max
~/app
∟ SessionStart: 1 open task here — Oathe is keeping track.
> fix the login guard
Task claimed. Guard rewritten — running the test.∟ saved · checkpoint
∟ You've reached your usage limit.
shez@mac app % oathe codex
🎉 Oathe just saved your session — 1 task still yours!open-source infra: github.com/oathe-ai/oathe
YOURS (completion not asserted)fix-login  restore the login guard · last: session ended (exit 0)
› continue fix-login from my Claude Code session
checkpoint restored · test rerun · done — asserted
said: "done" did: diff · test SETTLED ✓
auto save across your harnesses
PROJECT
oathe
LICENSE
Apache-2.0

Why aren't they running the company?

02:14computer turned off. 6h of work lost
09:31agent: "task complete". it's lying
11:02ran with prod credentials
14:45deduplication fails. client gets charged twice
17:20revoked key still works
+24dsilent failure. no alerts

What happens when the agent dies mid-task?

Left: an agent session. Right: the Oathe work record. Press kill.

kills survived: 0

The agent session

The Oathe work record

SETTLED ✓
work_claim
contract
attempt
effect 7d3a…
statements0
verdict
A replay of real tests: the agent was killed at six points, three times each. No work lost. No double effects. "Done" is verified. Evidence →

Every unit of work is an oathe.

Agents take actions, and just like with people, static checks are often insufficient. Agent reliability is a behavioral problem. We're solving it.

WorkClaim An obligation with one owner and one lifecycle. Outlives every process assigned to it. why it exists: the duplicate-session pileup
VerificationContract The standard the work is judged against — frozen at dispatch, before the agent runs. why it exists: our yield gate fired 28 times. wrong all 28
EffectReceipt Did the email send? Claim → attempt → receipt → reconcile. Unknown is a bounded state with a named resolver. why it exists: replay re-fired an external effect
ContextCompilation Context is compiled from company state with full provenance — every source, hash, and omission recorded. why it exists: seven uncoordinated injection paths
ScopeBound Tools, servers, model, budget — computed per claim from authority, enforced before the harness runs. why it exists: ambient config leaked into sessions
Settlement "Done" is a claim; only verification settles it. Replays can't settle twice. why it exists: an agent said done. it wasn't.

Scan a skill.

Any harness can publish a new skill to the company runtime. We audit and sandbox it first, so it can't break anything else.

200 free scans a month, no signup. Open Scan →

Oathe runs on the open standard.