Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

the agent’s paper trail

A security patch, fresh-context agents, confidence probes, and the interface of a useful agent reply.

Published Updated Read 3 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-09-11

the agent’s paper trail

An agent can now find a bug in a public database, lose the middle of a long task, carry too much of its own context, or write a reply that misreads the job. None of this is a new model announcement. It is the part that starts to matter after the demo: what happened, where it went wrong, and whether the final artefact fits the work.

1. Datasette puts the audit in the release notes

Datasette shipped security patches for both its alpha branch and its stable 0.65 line. The issue concerns public instances that mix public and private tables.

The useful detail is the working arrangement. Coding agents helped find problems, but one human wrote tests while another implemented fixes. Both reviewed every issue for nearly a week. This is less glamorous than a model finding a bug in the dark. It leaves something a maintainer can inspect when the fix arrives.

reading: the Datasette security release · Datasette’s release notes

2. Fresh context comes with a coordination bill

Subagents vs Agent Skills compares two ways to run reusable procedures. One keeps loading instructions into the main agent’s growing context. The other sends a subtask to a fresh-context agent with a defined input and output.

The paper reports better results for the latter on long tasks when the procedure has a clear contract. It also names the cost: more communication tokens. A fresh worker is not an escape hatch; somebody still has to specify the handoff well enough that the worker does not make up its own job.

reading: the paper on arXiv

3. A confidence signal at the moment of action

Do Agents Know When They Succeed? probes internal representations while an agent works through Bash, SQL, and Python tasks. The authors report that their monitor can estimate success without changing the prompt or paying for repeated rollouts.

That is a narrower claim than “the model knows when it is right.” It is a benchmark result, across three task types and three model families. Still, a signal that arrives before the final fluent answer is the right direction for systems that must decide when to stop, ask, or hand work back.

reading: the paper on arXiv

4. The last line is part of the interface

The i-have-adhd skill package tells coding agents to lead with the next action, use numbered steps, and lose generic closers. Its GitHub page shows 25,722 stars, which says more about appetite for a better reply shape than about any clinical outcome.

The package is small, but it names a neglected control surface. A reply that buries the next move under reassurance and explanation is annoying and makes a tool harder to use while somebody is already trying to steer a task.

reading: the repository

left on the table

  • LazyMind promises a self-evolving agent loop. The available material is a product claim, not enough evidence for a public dispatch.
  • OpenBiliClaw is a fresh coding-agent project, but the repository does not yet establish a change worth making into a story.
  • OpenMAIC appeared in Radar on 2 September. Same signal, no second lap.
agent-security agent-evaluation agent-workflows personal-ai