You pull the filing, scrape the page, extract the numbers — then start over for the next company.
One question runs an entire research team, phase by phase.
Twelve specialist agents split across five phases — plan, gather, verify, deliver, learn — turn a research question into a sourced report or slide deck.
Public repo · Python CLI · brings your own Anthropic API key
$ stratagem "Compare Arista and Juniper on competitive positioning"
▸ Running...
▸ research-planner [sonnet] planning research approach
▸ data-extractor [sonnet] extracting data
[scrape_url] → arista.com/investors
▸ financial-analyst [opus] analyzing financials
[download_sec_filing] → ANET-10-K.pdf
[parse_pdf] → ANET-10-K.pdf
▸ research-synthesizer [opus] synthesizing findings
▸ source-verifier [sonnet] verifying sources
▸ report-critic [sonnet] evaluating report quality
✓ report-critic done
[create_pptx] → Arista_vs_Juniper_Competitive_Analysis.pptx
━━━ DoneA control agent runs the whole pipeline
A `research-orchestrator` agent decomposes the question, then fans work out to twelve named subagents split across Plan, Execute, Quality, Deliver, and Learn — `research-planner`, `data-extractor`, `financial-analyst`, `research-synthesizer`, `source-verifier`, `report-critic`, and more, each pinned to Opus or Sonnet by role.
SEC filings come from EDGAR, not a scrape
`search_sec_filings` and `download_sec_filing` pull straight from SEC EDGAR. `parse_pdf`, `read_spreadsheet`, `read_pptx`, and `read_docx` extract text, tables, and images from whatever you hand it — cached per input so the same file isn't parsed twice in one session.
Three agents check the report before it ships
`plan-validator` checks for drift from the original question, `source-verifier` checks the claims against the sources, and `report-critic` grades the draft — all before `create_report` or `create_pptx` writes the final file.
Spins up a specialist when it hits one it doesn't have
`create_specialist` spawns a temporary agent to cover a capability gap found mid-plan. `after-action-analyst` then records what worked back into `.stratagem/` memory, so the next question in the same topic starts from what the last run learned.
- Tracing is off by default — set LANGSMITH_TRACING=true to opt in
- Every run's cache, filings, extractions, and reports stay in your project's .stratagem/ folder — nothing centralized
- SEC filings come from EDGAR directly, never scraped summaries
What does Stratagem actually produce?
A markdown, Word, HTML, or PowerPoint report via `create_report`, or a data workbook via `create_spreadsheet` — written to your project's `output/` or `.stratagem/reports/` folder.
Where do the SEC filings come from?
`search_sec_filings` and `download_sec_filing` pull straight from SEC EDGAR — not scraped summaries or a third-party cache.
Does it work outside Claude Code?
The MCP server (`stratagem.server`) is host-neutral — callable from Codex, Claude Code, or any MCP-compatible runtime. The `research-orchestrator` agent and its four skills are Claude Code-specific; Codex drives the same tools directly.
What stage is this at?
Beta. It's a public repo you can install and run today, but the agent set and phase pipeline are still moving — expect names and tool signatures to shift before a 1.0.
Every market research question turned into the same five-tool relay: pull the 10-K from EDGAR, scrape the competitor’s pricing page, extract the numbers from a PDF, run the ratios by hand, then write the memo — losing context at every handoff.
Stratagem is the strategy team I wanted instead of the relay: one orchestrator decomposes the question and fans it out to twelve specialist agents across plan, execute, quality, and delivery phases, gated by a validator, a source-verifier, and a critic before anything ships. It’s beta — the agent set and phase pipeline are still moving — but the orchestrator, the MCP tools, and the direct SEC EDGAR integration are real and running today.
Public repo · Python CLI · brings your own Anthropic API key