Build a weekly multi-agent research radar

Track a technical field with separate subagents for discovery, source quality, and skeptical review. The parent reconciles their findings into a weekly briefing.

The report appears in workflow output and history. Keep mutation tools denied on the research agent.

1. Configure web search

Choose one supported provider and verify it with an interactive agent before scheduling. See Web search for current provider fields and query differences.

The research agent needs web_search, web_fetch, http_request, spawn_subagent, and context tools. Create or edit an agent named research-radar and keep mutation tools denied.

2. Create the workflow

Create ~/.jazz/workflows/agent-harness-radar/WORKFLOW.md:

---
name: agent-harness-radar
description: "Weekly evidence briefing on open-source AI agent harnesses"
schedule: "0 8 * * 1"
agent: research-radar
autoApprove: low-risk
catchUpOnRestart: true
maxCatchUpAge: 172800
maxIterations: 70
maxCostUSD: 1.50
maxDurationMs: 900000
---

# Weekly agent-harness research radar

Find material published or released in the last seven days that changes how developers can build, evaluate, secure, or operate open-source AI agent harnesses.

Delegate three isolated investigations:

1. **Primary-source scout:** releases, changelogs, repositories, papers, and official engineering posts.
2. **Adoption scout:** concrete developer use, migration reports, benchmarks, and failure reports.
3. **Skeptic:** challenge novelty, publication dates, benchmark quality, and duplicated coverage.

Each investigator must return URLs, publication dates, one-sentence claims, and the evidence supporting those claims. Search current sources; do not answer from model memory.

Reconcile their results yourself. Drop items that are outside the date window, repeat another source, lack accessible evidence, or merely announce a feature without explaining its consequence.

Return:

## What changed

At most five items. For each: title, date, direct source URL, verified change, and why it matters.

## One claim worth testing

The most consequential claim, what evidence currently supports it, and a small reproducible test.

## Disagreements and uncertainty

Conflicting evidence, inaccessible sources, or conclusions the team could not verify.

## Quiet-week note

If nothing material changed, say so. Never pad the report to reach a target count.

Change the topic and source expectations to match the field you actually follow. A narrow decision-oriented question produces a better radar than “AI news.”

3. Test the exact scheduled policy

jazz workflow run agent-harness-radar --auto-approve
jazz workflow history agent-harness-radar

What you should see

Check that the report contains the requested sections and that each material claim links to a source published in the last seven days. When nothing relevant was published that week, it should say so.

4. Schedule it

jazz workflow schedule agent-harness-radar
jazz workflow scheduled

The report remains available in workflow history and scheduler logs. To receive it on your phone or in chat, configure a notification target and add deliver: <target-name> to the workflow frontmatter.

Read Starting runs, Scheduled runs, and Delegation.