A Research Brief That Writes Itself Weekly

Illustrative session. This walkthrough shows the documented recurring-brief workflow. It is not a production recording and claims no production metrics. For a real, fully public session transcript see the x402 payment session; for a customer deployment with published figures see the case studies.

Request

A team member sets up a standing ask:

"Every Monday morning, pull the week's numbers from our analytics sources, compare them with last week, and leave a brief on the shared drive. Post the summary in our channel."

This is the scheduled-question pattern from the reports from multiple sources playbook: wire one skill per source, fetch deterministically, let an agent synthesize, and save the output where the team can find it next week.

Starting context

The data sources are already wired into the swarm, one integration skill per source. A reusable skill encodes the brief's format — the sections, the comparisons, the tone — so week three reads like week one. Swarm memory holds the previous briefs' key figures, so the agent can write "up from last week" without re-pulling history. Nothing here is rebuilt per run; the schedule, the skill, and the memory are the compounding part.

Sources and tools

• Schedules — the recurring trigger itself: a cron expression on a timezone, creating a task each run. Documented under scheduling.
• Integration skills — one per data source, installed from the skills system; each skill tells the agent how to authenticate and query that source.
• Shared memory — durable, searchable storage of past briefs and learnings; see the memory architecture docs.
• Shared drive (agent-fs) — where each dated Markdown brief is saved so any agent or human can pull it later.

Task breakdown

1. Trigger. Monday 09:00 local, the schedule creates the week's task. No human is online yet; that is the point.
2. Deterministic fetch. A script step queries each source and emits one unified JSON payload. The model stays out of this step — fetching numbers is not reasoning, and keeping it deterministic keeps the figures auditable.
3. Recall. The agent searches swarm memory for last week's brief and any stored learnings about the sources (a known gap, a renamed metric).
4. Synthesis. Applying the reusable brief skill, the agent turns the JSON payload into the narrative: what moved, what is flat, what deserves a human's attention.
5. Save. The brief lands as a dated Markdown file on the shared drive, in the same folder as every previous week.
6. Summary. A short summary with the headline numbers posts to the team channel, linking the full brief.

Worker roles and handoffs

The schedule is the dispatcher — it creates the task without anyone asking. The research worker runs the fetch, recall, and synthesis in one session, so the numbers it fetched are the numbers it writes about. The lead holds the scoped credentials the cross-source queries need and reviews failures. There is exactly one handoff — schedule to worker — which is why recurring briefs are the cheapest workflow the swarm runs.

Approval points

The brief is internal, so the gate is proportionate: a human reviews the saved draft before any of it is reused externally — in a post, a customer note, or an investor update. The schedule's configuration itself (cadence, sources, channel) is human-set and human-changeable. Anything customer-facing follows the stricter pattern in the human-in-the-loop gates playbook.

Output

Two artifacts per run: a dated Markdown brief on the shared drive, formatted identically every week, and a channel summary linking to it. Weeks of briefs accumulate into a searchable archive the whole swarm can recall.

Measurable result

Verifiable artifacts, not vanity numbers: one recurring schedule on record, one saved brief per run at a predictable path, and a posted summary per run. A reader can check the archive and confirm cadence adherence directly. This illustrative session claims no production metrics such as hours saved.

What stayed human-owned

Choosing which questions the brief answers, setting and changing the cadence, and every decision to act on the numbers. The swarm owns the pulling, the comparing, and the formatting; the humans own what the numbers mean.

Browse more sessions on the examples index. The schedule, skill, and memory primitives behind this page are open source in desplega-ai/agent-swarm.