3 AI Agents, Zero Shared Context: Deep Research That Verifies Itself
| OWN THE EDGE |
| Aug 31, 2026 · AI & Automation for Builders |
|
|
Hi everyone,
Here's the thing about a single AI answering a research question: you're trusting one model's best guess. This week I built a system that refuses to do that. Three independent agents, three different models, zero shared context — each digging into the same question on its own. When they all arrive at the same fact through different paths, that's verification. Then a fourth agent fuses it, and a fifth saves every single claim and source to a live database as an audit trail.
I put real money on it — a full research pass on a real estate investment. Here's how the whole thing holds together.
|
|
🎥 Video Deep Dive
How I Built a 3-AI Agent Deep Research Workflow With Hermes Agent Bot Mode
|
|
The Build: Hermes just dropped Bot Mode — you build a specialized agent by giving it a name, a model, and specific skills and memory, then reuse it across any workflow. I used it to assemble a "deep research diamond." A workflow architect agent (I call him "Art") takes a project spec and breaks it into the agents needed: three scout agents running on DeepSeek-V4, GLM-5.3-Flash, and MiniMax-M3, researching independently with zero shared context. A GLM-5.3 fusion agent synthesizes what they found. And a Supabase agent persists every claim and source to a live database the whole time — not tacked on at the end, but a full participant.
The whole thing gets wrapped into a single deployable skill — a resumable diamond-pattern graph runnable via terminal, with mandatory citations and preserved attribution for disagreements.
|
|
🔧 Key Skills You'll Gain
|
Verification through independent paths. Three models, zero shared context, same question. When three different AIs reach the same fact through three different routes, you get real verification — not a lucky guess from one model.
|
|
The diamond fusion pattern. A lead researcher doesn't average results or pick a winner — it buckets where agents agree and where they clash. Agreement becomes a finding; disagreement gets preserved as a question, not smoothed away.
|
|
Bot Mode specialization. Define an agent as a profile — its own prompt, skills, MCP server, and memory, plus a self-improvement loop so it gets better as it learns. Then reuse it across workflows.
|
|
A workflow architect agent. Hand it a project spec; it decomposes the goal, defines the roles, sets system prompts and model choices per agent, and outputs a deployable skill — complete with a visual flow diagram.
|
|
Supabase as an audit trail. Every source and quote becomes an auditable database row. A verification gate checks the database directly instead of taking the model's word for it — and because you re-run over time, you can track trends, not just single snapshots.
|
|
|
📊 The Live Test: Real Estate With Real Money On The Line
I ran the full system on a Kitchener-Waterloo-Cambridge real estate investment question — 3-5 year horizon, risks and catalysts. The output: 91 claims backed by 74 citations, all persisted to Supabase.
The synthesis surfaced insights no single model would have flagged: a 50% condo market decline tied directly to federal study permit caps, a 10-quarter trend line, relative pricing vs. Toronto, and six catalysts driving the move. Most valuable were the four things no single model would have told you — the disagreement zones the fusion agent refused to bury.
|
|
Why It Matters: The default way people use AI research is a single model giving a confident answer with no way to check it. This workflow inverts that — verification is built into the architecture, not bolted on after. Independent scouts, a fusion agent that preserves disagreement, and a database that keeps every source auditable. The result is research you can defend: every claim traces back to a source, and every source is a queryable row. That's the difference between "the AI said it" and "here's the evidence."
|
|
📰 AI & Automation This Week
🔤 ANTHROPIC |
| The Anthropic Economic Index Connector |
Anthropic released a connector mapping real-world AI adoption data across industries — showing which sectors are actually deploying AI in production vs. just experimenting. Valuable signal for anyone building automation products or deciding where to focus. |
🔤 SEARCH ENGINE JOURNAL |
| Anthropic's Claude Can Now Watch A Video And Learn Your Job |
Claude can now watch screen recordings and learn workflows by observation — see a task performed once, replicate it. The industry is converging on skill permanence and multi-agent orchestration as the next frontier, the same patterns driving this week's build. |
🔤 PCMAG UK |
| Zapier Review: Boost Your Business With Easy Automation |
PCMag's latest review highlights how Zapier has evolved beyond simple triggers into a legitimate business automation layer — with AI-powered steps, multi-step logic, and enterprise features. The line between "no-code tool" and "operations platform" keeps blurring. |
|
|
|
|
Happy building,
Derek
|
|
How was this week's newsletter?
⭐️⭐️⭐️⭐️⭐️
Reply with feedback or more on YouTube
|
|
|
|