Urgent.News

What's breaking now, across thousands of outlets.

AI

Decide the Agent Seat by the Bundle You Can Carry Out

I would not park an agent on a free server until I can carry the whole run away. A friendly login and a warm machine can still hide the files, the tool transcript, and the prompt that produced a bad patch. If I cannot export those three things after one dry run, I treat that host as a sketchpad only. Would you trust a laptop you cannot image, just because the sticker on the lid says free? The…

When deciding which agent seat to use, the most crucial factor is whether you can export the essential files after a single run. If you cannot export the prompt file, tool-call log, workspace hashes, and a brief note of the blocked commands, that host is only suitable for temporary use as a sketchpad. You should not trust a host you cannot image just because it is labeled as "free."

The key questions are not which invoice looks smaller, but who holds the workspace after the process finishes, and who can destroy it without notifying you. Self-hosted boxes and paid isolated machines typically provide a disk that you can still access the next day, while shared free servers may simply disappear when the operator reclaims the seat.

Before selecting a host, you need to define the "exit bundle," which is the smallest folder that allows another machine to audit your decision-making process. This bundle should contain the prompt file, tool-call log, workspace hashes, and a short description of any commands you refused. If any of these components are only available within a web console, you do not yet have a defendable bundle.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

Understanding MCP Servers

The first time I heard about MCP, It seemed like something really complex. But it isn't. I remember the first mcp I used was the Figma MCP to replicate the design my fellow designer had created.

  • MCP stands for Multi-Modal Communication Protocol
  • Enables AI to interact with external tools and data sources
  • Host communicates with MCP server via client

Harness Engineering 101: How Coding Agents Actually Work

Take one model and give it 169 real bug-fixing tasks from SWE-bench Verified. Keep the weights, the tasks and the context window exactly the same.

  • Harness engineering optimizes model performance in tasks
  • Agent combines model with harness to complete tasks
  • Loop engineering wraps harness in additional loops

More from Thursday 24 September →