Broad discovery
Source lanes pull public signals — official pages, GitHub, releases, HN, feeds, packages, papers — without paid search calls.
Agent Radar · Markdown-first · GitHub Actions
Unattended research in Markdown
A Git-backed research loop for the AI-agent industry: collect public signals, screen candidates, write within bounds, and update weekly theses, watchlists, field notes, and storage implications.
$ python scripts/agent_radar.py brief
CORE FILES
-
radar.md -
agent-watchlist.md -
user-field-notes.md -
playbook.md -
storage-angle.md
radar.md Thesis: local agents win on evidence · Changed: sandbox snapshot noise · Open: who owns artifact storage
Markdown-first · GitHub Actionschecks before commit · Git as audit
One research loop
Markdown state, scheduled runner, two models
Agent Radar runs on GitHub Actions: Markdown holds durable research state, a deterministic runner scores and gates writes, and a cheaper model screens before a stronger model drafts.
Illustrative sample radar.md · sample week
Current thesis Local agents win on evidence
Coding agents that keep workspace artifacts local are easier to audit than chat-only hosted flows.
Changed Sandbox + snapshot noise
More products advertise sandboxes; few document snapshot retention or replay cost.
Open questions Who owns the artifact store?
When agents write logs, checkpoints, and playbooks, which storage path is intentional vs incidental?
Auditable state
What the pipeline leaves in Git
The runner updates weekly theses, the agent watchlist, field notes, and storage implications within fixed write bounds. When a candidate enters the watchlist or brief, the change remains reviewable in repository history.
- 01
Weekly thesis
Current judgment, what changed, and open questions for the Agent industry.
- 02
Watchlist
Mainstream and emerging agents with evidence labels — not a raw link dump.
- 03
Field notes
Real workflows and reusable patterns captured as they show up in the wild.
- 04
Storage implications
Where agents keep workspace state, sandboxes, logs, and replay — a side lens, not the whole product.
How the pipeline runs
How the pipeline runs
Score and dedupe
The runner scores items by lane reliability, relevance, adoption evidence, and novelty against the source cache before any model sees them.
Screen, then write narrowly
Discovery stays wide; promotion into the watchlist, weekly brief, or storage notes is bounded. The model may only update allowed files.
Check, then commit
Citation, quota, bilingual, and freshness checks run before commit. Failures stay visible in telemetry and issues — Git remains the audit log.
Ops and cadence
Ops and cadence
Scheduled on GitHub Actions
Daily, weekly, and monthly jobs run the same research loop unattended, with bilingual report output. Local runs are for debugging; Actions is the intended host.
Model suggests; runner decides
A lighter model screens; a stronger model drafts. The model may suggest updates; the runner scores, bounds writes, runs checks, and commits — labeling failures without blocking the whole loop.
FAQ
Who is this for?
People who want an auditable, scheduled research loop on the AI-agent industry — not a news feed and not a private note vault with no promotion rules.
Do I run it only on GitHub Actions?
Actions is the intended unattended host. The same runner can be invoked locally for debugging; the durable state remains the Markdown repo.
Do I need paid models or paid search?
Collection uses free public source lanes — no paid search. Bring your own model keys: a lighter model can screen, a stronger one can draft, so spend stays intentional.
How is this different from a Markdown notebook?
A notebook stores notes. Agent Radar runs collect → screen → synthesize → validate → commit with promote boundaries and checks, so Git history is the audit log.