For the next 25 minutes we're going to make exactly these choices for our discovery agent.

Today's stack: Sonnet 4.6 + Haiku 4.5. Two models, one cascade — cost reasoning in Part 5.

Plus — 12 transferable patterns & 7 other agents on the same stack.

Today's demo lives in YELLOW — cold outreach with a human checkpoint.

Pick one Green from your list. That's your first agent.

You'll get the slides · the speaker notes · the workflow JSONs · the autopilot prompts · the tutorials. Don't memorize anything. Reference. Pick your track this weekend.

Every input that drives the agent — voice, ICP, schemas, examples, exclusion lists — lives as a versioned editable file in Drive or Sheets. Not buried in workflow nodes.

// Claude Haiku 4.5 with web_search tool enabled
{
  "model": "claude-haiku-4-5",
  "tools": [{ "type": "web_search_20250305" }],
  "system": "[ICP keywords from icp.md]",
  "messages": [{
    "role": "user",
    "content": "Find founders publicly displaying ICP signals on
                HN, Reddit, Product Hunt this week. Return JSON:
                [{ person, signal_type, source_url,
                   evidence_quote, score 0-10 }]"
  }]
}

Plumbing. The unsexy stuff that separates "demo agent" from "agent that runs every day."

{
  "model": "claude-sonnet-4-6",
  "system": "[contents of voice.md] + [5 example emails]",
  "messages": [{
    "role": "user",
    "content": "Draft a customer-discovery email to {person}.
                Their signal: {evidence_quote}
                Source: {source_url}
                Their company context: {enriched_summary}

                Follow voice.md exactly. Interview ask, not pitch.
                Soft CTA. 80-110 words."
  }]
}

# Sophia's voice — for outbound writing

## Posture
- Direct, no fluff. Short sentences. Cut adverbs.
- Specific over abstract. Numbers over adjectives.
- Confident but not arrogant. Assume the reader is smart.

## Words I use
"architecture", "compounding", "leverage", "the math here is"

## Words I never use
"synergy", "unlock", "revolutionary", "I hope this finds you well"

## Email structure (cold)
1. One-line context: how I found you (specific)
2. Two-line observation: something specific about your situation
3. One question I'd ask
4. One sentence on what I do
5. Soft CTA: "worth 15 minutes Thursday?"

## Examples
[5 real emails — see voice-md-template handout]

The agent inherits your taste through examples, not adjectives. If you can't write 5 good examples, the agent can't either.

Subject: Discovery Pulse — May 8

Today:
  · 5 new drafts queued
  · 3 follow-ups ready
  · 1 reply yesterday from @kbrant

Cost today:  $0.21
Top signal:  pricing-frustration (n=2)
Your move:   approve before 5pm

date · leads_found · drafts ·
  cost_usd · avg_score ·
  top_signal · errors

2026-05-07 · 30 · 5 · 0.21 ·
  7.4 · pricing · 0
2026-05-08 · 28 · 5 · 0.19 ·
  6.9 · hiring  · 0

Trust gets built through visible track record. After two weeks of high approval rates, you can promote the agent to Stage 3.

Same architecture as Step 6. Different inputs. Once you've built one agent, you've built five.

To run · 4 free accounts: n8n.cloud · Anthropic ($5 credit) · Firecrawl (500 free) · Google. · To customize for your business · 4 things in your Sheet: voice.md (style) · icp_description (target) · signal_keywords (listen for) · subreddits (where they hang out). No redeploy.

The spine of every agent you'll build for the next 5 years. Tools change. Patterns don't.

Combine them. Recombine them. Tools change. Patterns don't.

Same architecture. Different trigger · search/extract · output. Ship a different one every week.

60–70% reduction. Same workflow, same output quality. The most underused cost lever in 2026. Free path follows the same logic — Llama 3.1 8B Instant for cheap, Llama 4 Scout for the heavy work.

Try Tier 0 free this week. No credit card. No install. Move to Tier 1 or 3 when you outgrow the trial.

Revisit local when API bill > $200/mo · PII workloads · predictable throughput.

Discovery Engine ships at Stage 2. Each row adds the guardrails of the row before. Graduate one stage at a time.

1 · Architect's Office Hours

5 free 30-min audits for BSW attendees.

You bring your bottleneck. I'll tell you what to automate, what to leave alone, and which 3 tools to use.

cal.com/sophia-stein/architect-audit-bsw

2 · Agentic Architect Blog

agenticarchitect.ai/blog

Weekly deep-dives on agentic architecture for lean founders. n8n templates · cost teardowns · what's actually working in 2026.

agenticarchitect.ai/blog

Repo: github.com/sudosoph/bsw26-agentic-workflows  ·  MIT · fork · ship yours