Agent Radar · Scheduled Markdown research

Agent news, with sources.

Public sources. Citation checks. Markdown briefs in Git.

PythonMarkdown-firstGitHub ActionsMIT v0.26.1 (Oct 2, 2026)

agent-radar 0.25.0 · captured output created
$ python scripts/agent_radar.py init
CREATED
  • radar.md
  • agent-watchlist.md
  • user-field-notes.md
  • playbook.md
  • storage-angle.md
radar.md in the repository · 2026-09-29 1. AI Agents are moving from chat and IDE autocomplete toward task-based execution.
Markdown-first · GitHub Actionschecks before commit · gaps labeled in Git

Research you can review in Git.

Markdown files hold the research. One model screens sources, then drafts; checks run before each commit, and gaps are labeled.

Quoted from radar.md · 2026-09-23 radar.md · main
Current thesis

Thesis 2

Coding agents are becoming the first high-frequency adoption path.

Changed

2026-08-21

Added signal: Cloudflare — MCP security updates (network detection, WriteGuard). … Evidence: Strong (official blog).

Open questions

MCP

Will MCP become the default tool integration layer? …

Every change stays in Git.

The runner updates four files within fixed write bounds, and each change stays reviewable in history.

  • Weekly thesis

    Current judgment, what changed, and open questions.

  • Watchlist

    Mainstream and emerging agents, with evidence labels.

  • Field notes

    Real workflows and reusable patterns.

  • Storage implications

    Workspaces, sandboxes, logs, and replay, as a side lens.

Oct 2, 2026

What’s new in v0.26.1.

  • The mainstream injector re-runs after the sweep replacement (before accountability/direction-quota), and the…
  • Daily results carrying updates but no day-block targets now fail fast with the actual shape problem instead o…
  • mainstream_auto_added accumulates across both injections

Get started

Python 3.10+ and only the standard library. Run it from a clone of the repository.

  1. Create the Markdown files

    python scripts/agent_radar.py init

FAQ

Does it need paid search?

No. Collection uses free public source lanes. You configure the model route; the pinned route runs GPT-5 Mini to screen and draft, about $2.6 a month against a $4 budget, with fallback models pinned in the workflow.

Where does it run?

Daily, weekly, and monthly jobs run on GitHub Actions and produce bilingual reports. If a Chinese half comes out thin, the runner regenerates it, or labels the report when it cannot. Local runs are available for debugging.

What if a check fails?

Most checks repair or label instead of refusing: a missing section gets an explicit gap line. Failures stay visible in telemetry, validate warnings, and issues; Git records what was published.

Agent Radar