Скачать или смотреть Deep Agents — Deep Research, Subagents & Middleware (Collier King | AIMUG Nov 2025)

Collier King (Machine Learning Engineer, Cloudflare) gives a practical, no-BS walkthrough of Deep Agents: when to use them, what they’re good at, the real engineering tradeoffs, and a live “deep research” demo that crawled 400 companies and ~244 pages of content to surface the top matches for themes from an industry keynote.

What you’ll learn

What Deep Agents are and how they differ from LangGraph-style deterministic flows.

A concrete deep-research use case: extracting themes, matching ~400 companies, validating with press releases, and ranking results.

How to design subagents (transcription → company matcher → PR validator → ranker) and organize long multi-step pipelines.

The middleware patterns that matter: remote storage (S3/R2), content truncation, logging, file-validation hooks and “before/after tool call” checks.

Practical pitfalls & fixes: schema drift, structured-output placement, token-efficiency, and keeping agents “on the rails” (explicit prompts + middleware checks).

Cost & performance example: Collier’s run burned ~11M tokens (~$80) and took several hours — tradeoffs to consider vs. LangGraph.

Tech takeaways (practical)

Use explicit schemas and examples in the prompt — don’t assume the model will follow an abstract schema.

Add middleware that blocks or raises errors on bad outputs (validation hooks) — treat the agent like a student you correct in real time.

For heavy, deterministic ETL-style work, LangGraph may be cheaper/faster; for open-ended planning and delegation, Deep Agents shine.

Logging + observability are non-negotiable for long-running jobs.

Chapters (approx — adjust to final video timings)

00:00 — Intro & overview (Collier King)

01:00 — When to use Deep Agents vs LangGraph

03:10 — Deep research use case: goals & dataset (400 companies)

06:00 — Subagents architecture: transcription, matching, validation, ranking

10:20 — Middleware: S3/R2, truncation, logging, validation hooks

14:15 — Common failure modes: schemas, skipping items, token shortcuts

18:30 — Cost & runtime: tokens, time, and tradeoffs vs LangGraph

21:00 — Demo log walkthrough & lessons learned

End — Q&A & closing

Resources & source

Slides, code and the experiment repo (linked in video description).

Transcript & source:

DeepAgents

Call to action
If this was useful — like, subscribe, and comment what you want Collier to deep-dive on next (middleware patterns, structured outputs, or cost-optimization).