Deep Research Team¶

Comprehensive web research with source collection, synthesis, and editorial review. Operates in workspace mode -- each run gets its own dated directory with structured output.

Orchestrator¶

The orchestrator has restricted tools (read,find,ls,grep via team-prompt.md frontmatter) and cannot fetch URLs or run commands directly. It must delegate all work through the subagent pipeline. The body of team-prompt.md gives it detailed context about the team's agents, workflows, and constraints.

Agents¶

scout¶

Field	Value
Tools	bash, read, ls
Skills	tavily

Searches the web via the Tavily API for high-quality sources on a research topic. Returns a numbered list of URLs with titles and relevance notes. Formulates multiple search queries to cover different angles.

collector¶

Field	Value
Tools	bash, write, read, ls
Skills	playwright

Fetches a single URL via headless browser (playwright-cli), strips boilerplate and ads, and saves cleaned content to sources/<slug>.md with YAML frontmatter. Takes a screenshot to screenshots/<slug>.png. Deployed in parallel -- one instance per URL from the scout's list. Each instance gets a unique collector number to avoid browser session collisions.

writer¶

Field	Value
Tools	read, find, ls, write
Skills	none (`no-skills: true`)

Reads all files in sources/ and synthesizes them into a structured research report with inline source references. Saves the draft to drafts/report.md. Follows a strict template: title, executive summary, subsections with citations, and a sources list with screenshot references.

editor¶

Field	Value
Tools	read, find, ls, write
Skills	none (`no-skills: true`)

Reviews the draft against source files for accuracy, completeness, and structure. Verifies citations, checks screenshot references, and produces the final report at report.md in the workspace root.

Workflow¶

The standard pipeline runs four steps:

scout (single) -- searches for sources on the topic
collector (parallel) -- fetches and cleans each source URL, saves to sources/ and screenshots/
writer (single) -- synthesizes all sources into drafts/report.md
editor (single) -- reviews and produces final report.md

Workspace Structure¶

Each run creates a dated directory under workspaces/deepresearch/:

workspaces/deepresearch/2026-04-12-141259/
├── sources/          # Cleaned source files (markdown + YAML frontmatter)
├── screenshots/      # Page screenshots (PNG)
├── drafts/           # Intermediate report draft
├── sessions/         # Session logs (orchestrator + all subagents)
└── report.md         # Final deliverable

Configuration¶

All orchestrator configuration is in team-prompt.md YAML frontmatter:

---
name: Deep Research
description: Search, collect, synthesize, report. Each run gets its own workspace.
tools: read, find, ls, grep
---

The body of team-prompt.md provides the orchestrator's system prompt with team context and workflow instructions.

File	Purpose
`workspace.conf`	Lists subdirectories to pre-create (`sources`, `screenshots`, `drafts`, `sessions`)
`team-prompt.md`	Orchestrator config (frontmatter) and system prompt (body)

Usage¶

# Start a new research run
p deepresearch "What are the latest developments in WebTransport protocol?"

# List past workspaces
p deepresearch --list

# Resume the most recent workspace
p deepresearch --resume

# Resume a specific workspace by prefix
p deepresearch --resume 2026-04-12

Running Evals¶

# Quick smoke test (4 prompts)
./evals/run-eval.sh deepresearch evals/deepresearch-short.txt

# Comprehensive suite (20+ prompts)
./evals/run-eval.sh deepresearch evals/deepresearch-long.txt

# Results saved to evals/results/deepresearch/<eval-name>/<timestamp>/