SEAHORSE: What Actually Makes an Agent Code Well

Posted by:

|

On:

|

SEAHORSE: What Actually Makes an Agent Code Well

An agent that writes code well is not primarily an agent that writes good code. It is an agent that can work alongside other agents on one tree without corrupting it, find out what is already true before acting, hold a few beliefs that survive a large codebase, and run on a harness that gives it those primitives.

Abstract

SEAHORSE takes those four in order. ORCA is a lease-based orchestrator for concurrent multi-agent editing: the correct unit of mutual exclusion is the file during application, not the task. A structured design corpus — a “bible” of single-owner optics with executable verification blocks — is then shown to assert shape, not liveness: every invariant can pass while the system it describes does nothing at all.

On 2026-08-04 that was not hypothetical — four headline features were dead simultaneously, one of them for eight weeks, with a fully green gate. Four premises about belief separate an agent that makes progress from one that stalls. The sea-themed name is deliberate: the tooling already sails under ORCA and MERMAID. v1.0, 5 August 2026.

This page is the public summary. The full paper (v1.0) is linked below as a PDF.

The implementation is part of Building Jarvis, an open series on persistent agents. Follow the work and contribute at github.com/globalcaos/tinkerclaw.

Read the paper

First page of SEAHORSE v1.0

📄 Read the full paper (PDF) →

11 pages · ORCA leases, the bible, four premises, the harness

Everything on the site, by topic or date: Browse the Tinker Zone

Built with TinkerClaw. Every tool and experiment on this site runs on TinkerClaw, my open-source fork of OpenClaw. TinkerClaw on GitHub · What we changed in our OpenClaw fork