Install in one command
digitorn install hub://digitorn/deep-researchWhat you get
- 5-agent pipeline with citations on every claim
- Fact-checker at temperature 0 to catch hallucinations
- Mixed-provider routing (Anthropic + DeepSeek)
- Working memory for cross-turn dedup
Credentials this template uses
Click each one for the credential setup command, scope guidance, and provider details.
The full YAML, ready to fork
app.yaml
1app:2 app_id: deep-research3 name: "Deep research"4 version: "1.0.0"5 description: "5-agent research pipeline with citations."6 category: "research"78modules:9 web:10 config:11 search_backend: duckduckgo12 memory:13 config:14 working_memory: true15 todo_list: true16 agent_spawn: {}1718execution:19 mode: one_shot20 entry_agent: coordinator2122agents:23 - id: coordinator24 role: coordinator25 modules: [{agent_spawn: [Agent]}, {memory: [set_goal, remember]}]26 brain:27 provider: anthropic28 model: claude-sonnet-4-629 credential:30 ref: anthropic_main31 scope: per_user32 provider: anthropic33 temperature: 0.234 system_prompt: |35 Decompose the question into 2-4 angles. Spawn researchers in parallel36 via Agent(specialist='researcher', prompt='...') plus a fact_checker.37 Wait on all of them with Agent(agent_ids=[...]). Then call writer,38 then editor, sequentially.3940 - id: researcher41 role: specialist42 modules:43 - {web: [search, fetch]}44 - {memory: [remember]}45 brain:46 provider: deepseek47 model: deepseek-chat48 credential:49 ref: deepseek_main50 scope: per_user51 provider: deepseek52 temperature: 0.25354 - id: fact_checker55 role: specialist56 modules:57 - {web: [search]}58 brain:59 provider: deepseek60 model: deepseek-chat61 credential: deepseek_main62 temperature: 0.06364 - id: writer65 role: specialist66 brain:67 provider: anthropic68 model: claude-sonnet-4-669 credential: anthropic_main70 temperature: 0.47172 - id: editor73 role: specialist74 brain:75 provider: anthropic76 model: claude-sonnet-4-677 credential: anthropic_main78 temperature: 0.0How it works
researcher pool with web tools only
Two researcher specialists run in parallel via the Agent tool. Each calls web.search and web.fetch on its assigned angle, then writes findings into shared memory.
fact-checker at temperature 0
The fact-checker has a narrower toolset (web.search only) and runs at zero temperature so it answers verification queries deterministically. Outputs a yes/no/citation triple per claim.
writer + editor as final stage
The writer takes the merged findings and produces the synthesis on Sonnet. The editor passes over it for terseness and flow at zero temperature. Both run sequentially, only after the parallel research stage finishes.
Read the deep dive
How we cut our coding agent's bill by 60% with model routing
Read article
Newsletter
Get the next post in your inbox.
Engineering notes from the Digitorn team. No marketing, no launch announcements, no "10 prompts that will change your life". Just the things we write that we'd want to read.
Other templates
developerStarter coder/templates/starter-coderdeveloperClaude Code clone/templates/claude-code-clonemessagingSlack helper/templates/slack-helperdeveloperPull request reviewer/templates/pr-revieweropsWeekly digest reporter/templates/weekly-digestknowledgeDocumentation generator/templates/doc-writercreativeLive React app builder/templates/react-buildermessagingVoice assistant/templates/voice-assistantopsScheduled data pipeline/templates/data-pipeline