Playbook · Published under Adrian Verdan · English
Orchestrating AI Agents
Learn when one model is enough and when an additional agent earns its place. A practical, dated playbook for designing and evaluating multi-agent workflows — with reviewer patterns, routing methods, research verification, cost controls and copy-paste prompts.
10 chapters
Interactive cost calculator
2 XLSX workbooks
Prompt pack
HTML bundle + PDF
Data as of September 2026
Language: English
€12.90
Final price · VAT exemption under section 19 of the German VAT Act
For consumers in the EU · PayPal · Immediate download available
Buying for a business or outside the supported EU territories? Buy on Gumroad ↗. Gumroad is the seller there and shows any VAT and the final price before purchase.
Created with AI assistance and editorially revised. Editorial responsibility: Adrian Verdan (Dargel Solutions). Transparency notice and details inside the product.
Film · 0:40 · Instrumental music, no voice-over
What the film shows
0:00 Three agents. Three answers. Zero decisions. · Agent A: Done. Rewrote the login flow. · Agent B: Done. Also rewrote the login flow. · Agent C: Reverted both. Mine is cleaner. · Agent A: Reverting your revert. · Agent B: Fine. Starting over from scratch. · Agent C: Who signs off on this? · Agent A: Not me. · NO ONE IN CHARGE
0:09 Not more agents. Clear roles.
0:12 Orchestrating AI Agents · 2026 edition · Know when one model is enough. · And when a second agent earns its place.
0:16 One writes. One checks. You decide. · Builder holds the pen · Reviewer read-only · reports · You make the call · Chapter 5 · the reviewer pattern, hands-on
0:22 Seven patterns. Most tasks need none. · Chapter 4 · the honest default
0:26 Check every prompt before it runs. · Security check, Chapter 10 · flag 4 of 8
0:30 Estimate the cost before you commit. · Example values · Ch. 8
0:34 Many agent voices. One human approval. · Playbook · 10 chapters · English
The motif
Many agent voices — one human approval.
Four agent lanes — plan, build, review, research — converge on a single human approval gate. That is the whole method, and it is the score motif the book is designed around.
From the book’s design language: notes are agent turns, outlined notes are review passes, and every strand ends at the amber human-approval ring.
What's inside
Ten chapters, from model choice to workflow.
Why more than one agent. The principle, a maturity ladder, and an honest account of when one model is enough.
The 2026 model landscape. Claude Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash and Grok 4.6: dated model references, access limits and migration checks. No default winner.
Open & self-hosted models. Llama, Qwen, DeepSeek, Mistral, gpt-oss: a decision tree for when local is genuinely worth it (hint: usually privacy, not cost).
Orchestration patterns. Reviewer, planner-executor, router, ensemble, debate, pipeline, orchestrator-worker: what each is for, and what it costs.
The reviewer pattern, hands-on. Reproduce a builder plus an independent read-only reviewer (Claude Code + Codex-style) step by step, including conflict resolution.
Tooling & orchestration stack. Framework choices, pending tool calls, cancellation and recovery. Use a small evaluation set to decide how much orchestration the task needs.
Deep research in your workflow. How agentic research works, which tools fit, and how to verify it before you build on it.
Cost & quality. A cost model, an interactive calculator, and evaluation-led recipes per task type.
Safety, security & governance. Prompt injection between agents, least privilege, and human-in-the-loop gates.
Start now. Six prompt templates for five workflows: security triage, planning, setup, review and two-stage research.
Real excerpts
This is what the product looks like inside.
Orchestrating AI AgentsChapter 1
When orchestration does not pay off
Start by testing whether orchestration adds value. Three situations where adding agents can cost more than they add:
The task has many interdependencies. When subtasks continually need shared information, handoffs can cost more than they add. Test your task against a single-agent baseline rather than treating coding as categorically unsuitable.
Handoffs lose context. In practice, agents “suffered from lost context at each handoff and spent more tokens coordinating than executing” (Anthropic, “When to use multi-agent systems,” January 2026). More agents means more places where information can be dropped or distorted.
The task is not well-defined enough to decompose. A fuzzy problem does not become sharper by splitting it across two models. Fix the problem statement first.
Excerpt from the playbook, chapter 1 — shown with full context in the product.
Orchestrating AI AgentsChapter 8 — Live demo
Multi-agent cost calculator
Pre-filled with the playbook’s documented defaults. Change any value to model your own workload; prices are USD per million tokens (MTok).
Enter valid non-negative values. Token counts, tasks and passes must be whole numbers of at least 1.
$0.29
Multi-agent cost / task
$9.00
Extra cost / month (reviewer)
45.0 %
Reviewer uplift vs. single
Single-agent cost per task = (inTok × builder-in + outTok × builder-out) / 1,000,000. Reviewer cost per task = passes × ((inTok + outTok) × reviewer-in + 0.5 × outTok × reviewer-out) / 1,000,000 — the reviewer reads the output plus the context and writes about half the output as a findings report. Multi-agent = single + reviewer; monthly totals multiply by tasks per month. Token inputs are raw counts; prices are USD per million tokens.
The calculator uses editable example rates, not a live model-price feed. Enter the current rates for your chosen builder and reviewer before budgeting; include cache, tool and retry costs separately. Model-specific long-context thresholds and service modes require separate pricing; the calculator is a flat-rate estimate.
Excerpt from the playbook, chapter 8 — the full version shows all seven metrics, the formulas and the levers that matter more than model choice.
Before you buy
An honest promise.
A separate review pass can reduce missed errors when the exact builder-reviewer pairing performs well on representative evaluations; it can also miss defects or add false positives. This guide gives you a method and a dated snapshot, not a guarantee.
Verify current model names, capabilities and prices on the provider’s official page before relying on them. The September 2026 update dates its model references and separates access, API compatibility and example costs. Provider limits and prices can change.
Full prompt pack (.md): six templates for five workflows — security triage, planning, setup, review and two-stage research
Frequently asked
Quick answers.
What exactly is included?
Ten chapters as an offline HTML bundle plus a tagged PDF, an interactive multi-agent cost calculator, a model-selection matrix (XLSX), an orchestration setup checklist (XLSX), and the full prompt pack as a plain-text file. Everything works offline.
Do I need a machine-learning background?
No. The playbook is written for solo founders and indie developers who already work with an agent like Claude Code, Codex CLI or Cursor and want a coordinated setup without a machine-learning background or unnecessary complexity.
Does adding a second model guarantee better results?
No. A separate review pass can reduce missed errors when the exact builder-reviewer pairing performs well on representative evaluations; it can also miss defects or add false positives. The guide gives you a method and a dated snapshot, not a guarantee — and it is honest about when one model is enough.
€12.90
Final price · VAT exemption under section 19 of the German VAT Act
For consumers in the EU · PayPal · Immediate download available
Buying for a business or outside the supported EU territories? Buy on Gumroad ↗. Gumroad is the seller there and shows any VAT and the final price before purchase.
Part of the AI-First series
The AI-First Operating Playbook is the business-wide operating system this deep dive plugs into: automation blueprint, function-level playbooks, governance and ROI. Orchestrating AI Agents is the hands-on multi-agent companion behind its chapter 5.
And if you want the operating half rather than the design half: Nobody Opens a Passing Check is a logged field report from thirty-nine days of running a two-agent team, read on day 39 of 90 (2026-08-25). It presumes a team exists and is about keeping it honest while it runs.
