What Is Oathe? An Open-Source Runtime for Agent Harnesses

Today, agent work belongs to sessions. A session dies and its unfinished work becomes your problem. A context window compacts and the agent reconstructs the job from fragments and a self-reflection prompt. A provider rate-limits you mid-refactor and switching harnesses means pasting a transcript and re-explaining everything. An agent says “done,” and that statement quietly becomes reality.

The problem is that AI is built to make you happy. Saying “done” and being “done” are two very different claims. When that touches client delivery or worse payment processing… how do you know the models did what they SAID they did?

That’s what oathe answers. Agents must take an oathe so their own work can be verified and checked for the human they are responsible for.

Oathe moves that responsibility out of the session. It’s an open-source runtime that sits underneath your harnesses — Claude Code, Codex, OpenClaw, and friends keep their own prompts, tools, and UX — and keeps the work itself durable, transferable, and verified.

What Oathe Is

Oathe lets personal AI Agents hand off and verify each other’s work. A task moves between harnesses as a context bundle — Claude Code to Codex, OpenClaw to Hermes, Claude Code to Microsoft Scout. Completion is never taken on faith: Oathe post-processes the agent’s trajectory against the task intent, and a non-author verifies before work settles.

Why does verification matter this much? Because models are getting too good at being wrong and confident. The new agent shouldn’t have to take the old agent’s word for anything — and with Oathe, it doesn’t.

How a Handoff Works

You’re mid-refactor in Claude Code when it hits a usage limit. Open Codex in the same folder and it picks up the same task: what’s claimed, what’s done, what’s left, and a test that decides whether it’s finished. No copy and paste, no re-explaining.

task claimed
  → attempt 1 (Claude Code) works in a scoped workspace
  → checkpoint · checkpoint · checkpoint
  → attempt 1 dies (crash, kill, rate limit, closed laptop)
  → task survives, still owned
  → attempt 2 (Codex) gets a compiled bundle + a checkpoint
  → attempt 2 inspects the diff and reruns the test before trusting anything
  → work completes; a non-author verifies the trajectory
  → only then is the task settled

LLMs are grounded in attention — it’s their underlying design principle. Model context is a cache; it gets compacted, truncated, and lost. Oathe compiles what the next attempt needs from saved state, so an agent remembers what the task actually was before making changes or decisions that could lead to negative repercussions.

No Agent Grades Its Own Homework

“Done” is a claim, not a fact. Oathe treats it that way: completion is asserted, never self-settled. A verification lane that didn’t write any code reads the agent’s own recorded session traces and judges whether the agent did what it said — not just whether the logs look plausible. It’s not an agent reading logs; the trajectory is post-processed against the task intent. You can even ask a different harness to run the verification.

We’ve written before about why agent reliability is a behavioral problem — this is our runtime answer to it.

Durable By Design

If an agent is rudely interrupted — network drop, power cut, rate limit, closed laptop — the task survives, still owned, and the next attempt picks right back up. No tokens wasted re-deriving what already happened.

Consequential actions get the same treatment: side effects carry durable receipts, so a deploy or a payment that already happened doesn’t happen again on a retry. The definition of done is established before work starts, so it can’t quietly drift.

The Ecosystem

Integration is everything. Oathe runs on the open standards — MCP, A2A, and AP2 — and treats your harness as a first-class citizen rather than something to replace.

Oathe Scan is our audit engine: it runs third-party skills in a sandbox and checks their behavior before they enter the runtime. Check any skill before you install it — we’ve audited over 1,600 of them.

Open Source Because It Has To Be

A trust layer between everyone’s agents can’t be a black box. All agents should be able to trust each other’s work and hand off context bundles seamlessly — that only works on an open substrate. The runtime is Apache-2.0 at github.com/oathe-ai/oathe, alongside the Open Threat Classification and the Oathe Scan MCP.

Get Started

npm install -g @oathe/oathe

cd your-project
oathe init          # local setup, installed harnesses onboarded
oathe claude        # a normal interactive session — the board shows

Ctrl-C anytime. Open oathe codex in the same folder — same board, same claims, other harness. Then oathe claim, oathe done, and oathe verify when you want the loop made explicit.

Questions, or a harness you want supported? The repo and Discord are open.