Agent Harness 不是 while 循环：omp²《Harness Playbook》
Before adding another loop, decide who owns state, who can stop the work, and where one run ends.
← All editions Editorial preview
The next phase of agentslives beyond the chat box
From authoritative harness state to takeover-ready interfaces and real deployment costs: eight reads that reconsider what an agent actually is.
From the editor
Editor’s note
Edited by AiringMost of this week’s saved pieces move past “how much stronger is the model?” and ask what an agent needs as a working system: authoritative state, bounded execution, interfaces people can take over, and costs that shape the product. Their answers conflict, especially on how much history to keep. That tension is why they belong together.
Lead story
01 / X 01 / State and control
Reliability means preserving verified progress
More context is not automatically better, and neither is less. The real question is which state must persist and which attempts are worth keeping.
After 70 Coding-Agent Rounds: Harness-of-Harness
After 70 rounds, the hard part is not generating more code. It is keeping the next round from erasing work that already passed review.
SKILLSTATE: Keep Only Current State
Keeping only current state and the latest observation is a clean rebuttal to the idea that more context is always better.
AGI Hunt · Chain-of-Experience: Don't Rush to Compress Attempt History
The counterpoint: feedback embedded in attempt history may itself be capability. The choice is not more or less, but what survives.
02 / Working interfaces
Put delegation, observation, and takeover in one place
If an agent keeps working, a chat transcript cannot remain the product’s only container.
Huabu：从线性对话到二维协作的 Agent 新桌面
A two-dimensional canvas turns delegation into a visible spatial relationship across collecting, organizing, and executing.
Designing Grok Bot for a world of persistent agents
Once a bot becomes persistent, the entry point is no longer a conversation but a recoverable unit of work that people can take over.
03 / Scale and fieldwork
Beyond compute, there is the cost of the organization
Infrastructure determines how many instances can run. Fieldwork determines whether any of them create value.
Obscura · Open-Source Headless Browser for Agents
A lighter browser is not a benchmark trophy. At hundreds of concurrent instances, 30 MB becomes a product boundary.
My FDE
Reducing ten minutes to ten seconds is not the same as creating value. Agents still have to meet real data, permissions, and people.