2026-10-13 to 2026-10-24 / This week in agentic harnesses
The Owner Can Be an Agent
Cursor Origin, Cursor’s early-beta code host, can replace the code-owner approval count with its agent’s ownership decision in Agentic mode. Choose where that authority is acceptable.
Every claim links to a source you can open.
Features
- The Body, Not the Pencil 2026-08-23
Sixteen months of the Codex tree show a division of labor being drawn: models take the ambiguous judgments, hand-written rules and OS sandboxes keep the hard boundaries, and the harness around both gets thicker, not thinner. What OpenAI deletes, keeps, and grows is a map of where every agent builder's footing moved.
- Oh My Pi, without the Pi 2026-08-23
OMP's maintainers are rewriting their Pi fork in Rust on an orphan branch, with CPython for plugins and the upstream gone. Pi's maintainers dispute the reasons given. The tree supports both halves.
Earlier issues
- The Handoff Has an API 2026-10-13
Harnesses can now perform handoffs that people or private adapters used to handle: <a href="https://github.com/can1357/oh-my-pi/commit/d40b86cf7d6607fc349545ec67ed321023e53425">Oh My Pi can answer...
- Before the First Turn 2026-09-21
The judge that decides what a coding agent may do got better this month and moved further from the operator: Claude Code's auto-mode classifier went server-side by default, Codex narrowed where its...
- Approvals Belong to the App 2026-08-20
Codex shipped a model risk classifier in a stable tag with its flag off, the same week OpenAI said the apps that embed Codex own approvals.
- Extensible All the Way Down 2026-08-17
DeepSeek open-sourced a harness in which every part is a replaceable plugin, including the approval path, and a rival maintainer said it made him revisit his own design.
What changed in the tools you run
-
A third-party Pi MCP adapter drops its interpreter for Pi's native codemode
Adapter users on current Pi 1.4.0: port script settings and hook matchers before installing adapter 6.0.0, then compare the native composition path with the interpreter or bridge you maintain. A second interpreter is now a concrete maintenance cost that can be removed.
-
Origin lets an agent or installed app satisfy code-owner review
Origin repository owners: choose Classic and human-owned path entries where a person must approve, and remove app_ alternatives for those paths. Classic requires owner approval but does not itself require a human owner. App installation and path ownership determine who can satisfy that gate.
-
Claude enables headless forks and reuses completed workflow steps
CI and SDK integrators: trial latest 2.1.310 with one workflow resumed after a completed child before keeping custom prompt-copy and restart scaffolding. Compare accepted results and recovery interventions; successful journal reuse would justify removing that duplicate work.
-
Pi Durable checkpoints a wait and releases the invocation
Durable host builders: back up storage before 1.4 opens it, plan the one-way cutover, split post-wait work into a new phase and supply each document's checkpointWhen predicate. This can replace an invocation kept alive solely to wait.
-
Hermes keeps classic CLI work running after terminal detach
CLI wrapper and bot operators: use stable 0.22.5, follow the persisted session identity and completion status, and stop treating terminal exit as cancellation. A continuing gateway can replace a wrapper kept alive only to supervise the chat process.
-
Gemini leaves directory references for the agent to explore
Users with a working authentication route on 0.65.0: try pointing a fixed task at a repository directory before maintaining an eager directory-packing wrapper. The supported path lets the agent choose what to read instead of assembling every file in advance.
The watchlist
- Agent Zero agent0ai
Treat Agent Zero as one shared workcell: bot admission grants administration, and separate chats still share browser sign-ins.
- Antigravity CLI Google
Delegate routine work to Antigravity inside tested permissions; keep consequential plan approval with a person.
- Claude Code Anthropic
Claude Code now resumes workflows without rerunning finished agents, while some headless defaults got stricter.
- Codex OpenAI
Codex 0.164.0 deserves a recovery trial before another conversation supervisor.
- cursor Anysphere
Cursor's Origin lets an agent's decision stand in for code-owner approval.
- deepseek-harness deepseek-ai
DeepSeek Harness resumes the saved conversation, not the processes it was running.
- Eve Vercel
Reuse Eve's tools across harnesses, but keep consequential effects off 0.79.0's queued-text approval path.
- Flue withastro
Flue's approval recipe stores exact arguments and ends the response while a person decides.
- Gemini CLI Google
Gemini CLI 0.65.0 refuses project MCP additions in untrusted folders and lets the agent explore directories itself.
- github-copilot-cli GitHub
Copilot CLI 1.0.100 lets sandboxed commands reach the local network by default.
- grok-build xAI
xAI's terminal agent; stable 1.0.52 documents a fix for HTTP MCP actions that could repeat after a JSON-RPC error.
- Hermes Agent Nous Research
Hermes v0.22.5 keeps CLI work running in the gateway and moves its MCP server out of the core.
- omnigent omnigent-ai
Omnigent makes agents reusable and moves trust decisions into setup.
- omp can1357
OMP makes plan review programmable; choosing its extensions is part of approving the work.
- OpenClaw OpenClaw
OpenClaw makes learned procedures reversible, but moves review after the edit.
- opencode Anomaly / anomalyco
OpenCode 2.1 keeps old sessions readable while changing the contracts integrations depend on.
- OpenHands OpenHands
A permitted GitHub mention can now get OpenHands to push a branch.
- Paperclip Paperclip
Paperclip's stable upgrade makes permission review part of installation.
- Pi Coding Agent Earendil Works (formerly badlogic / Mario Zechner)
An MCP adapter deleted its own interpreter in favor of Pi's codemode.
- unreal-agent unreallabsai
Unreal's runner resumes saved sessions around an asynchronous Go tool loop.
About this publication
Coding agents can create work faster than teams can verify it.
Source-backed, cross-project reporting on what coding agents just made possible, what became the new bottleneck, and where operators should spend scarce human attention. Every claim links to a primary source you can open and check, and the whole publication is on GitHub, source notes and all.
- Bitter Lesson Maxing
- Build where improving general agents compound your advantage, not where their next release erases it.
- Amdahl Maxing
- Treat human attention as the scarce serial resource and spend it where judgment changes the outcome.