Companion page Updated quarterly AI Agents for Everyone
The book is ink on paper - or pixels on a Kindle screen - and once it's printed it stays still. The field doesn't. This is where the model lineup, indicative pricing, MCP ecosystem and any errata get refreshed without reprinting the book. Public page - no login, no gate.
Edition October 2026 · 2nd ed. Last verified 10 October 2026
Old Forge · Field Manuals
AI Agents
for everyone.
From ChatGPT to autonomous
systems that actually work
Buy the book
The book references this page in the Introduction, the Pace and Evolution section, every pricing-table footnote, and the Resource Directory. Every code snippet from the book is also reproduced, copy-paste ready, in the code appendix.
Names rotate every few months; the tiers are stable. The book argues you should pick a tier first - based on the work the agent has to do - and only then pick a specific model. The names below are accurate as of October 2026; verify at the provider before relying on any of them in production.
Tier 1 · Frontier-Opus class
Best-in-class reasoning, tool use and long-context coherence. Use when the work genuinely needs it - research agents, multi-step planning, code at the edge of the model's ability.
Tier 2 · Frontier-Sonnet class
The default choice for most production agentic work. Strong reasoning, fast enough for interactive use, far cheaper than Tier 1.
Tier 3 · Mid
For the boring half of your pipeline - intent classification, routing, summary drafts, retrieval re-ranking. Cheap enough to run at volume; capable enough to do the work well.
Tier 4 · Small / on-device
Embedding, simple extraction, edge inference. Pennies per million tokens, or free if you're running open weights on your own kit.
Names verified at provider sites on 10 October 2026. Anthropic, OpenAI, Google, xAI and Mistral all rotate model names within a tier without warning; treat this list as a starting point, not a contract. Retired since the first edition: GPT-4, GPT-4-turbo, GPT-3.5-turbo, the o-series, and the Claude 3/3.5 family.
Bands, not point values, because providers move pricing every few months. Conversions assume £1 ≈ US$1.27. Always check the live provider page below before quoting numbers to a client.
| Tier | Input £ / 1M | Output £ / 1M |
|---|---|---|
| Tier 1 · Frontier-Opus | £3 – £8 | £16 – £40 |
| Tier 2 · Frontier-Sonnet | £0.60 – £1.60 | £3 – £8 |
| Tier 3 · Mid | £0.08 – £0.25 | £0.40 – £1.20 |
| Tier 4 · Small / on-device | £0.02 – £0.16 | £0.10 – £0.80 |
Prompt caching, batch APIs and committed-use discounts can move the effective rate significantly below the band on Tier 1 and Tier 2 - see the book's pricing chapter for the conditions under which each one is worth claiming.
Two years ago, 32K tokens was the working ceiling. As of October 2026, every Tier 1 and Tier 2 frontier model ships a ~1M-token window (OpenAI's GPT-6 line runs to about 1.05M; xAI's Grok 4.7 offers 500K). The practical implication is that whole codebases, full regulatory packs, and the back-history of a customer account now fit in a single call - though the cost of using all of it on every request remains the constraint, not the window itself.
The book's guidance still holds: retrieval beats long context on cost almost every time, but long context lets you stop worrying about chunking on the cases where it doesn't.
Model Context Protocol has gone from a Claude-only experiment to a cross-vendor de facto standard in roughly eighteen months. The book argues integration time has collapsed from weeks to days for mainstream tools - this is the shape of that, today.
Shipping · first-party
In beta or limited · first-party
Community-only · vendor not yet committed
The base list was audited April 2026; the three low-code additions above were verified October 2026, alongside the MCP specification 2026-07-28 (stateless core, hardened auth) and the A2A agent protocol under the Linux Foundation. Community servers vary wildly in quality. The book's chapter on MCP explains the questions to ask before relying on one - auth model, rate-limit handling, schema fidelity, who maintains it.
Corrections to the April 2026 first edition, all fixed in the October 2026 second edition. Filed 10 October 2026.
| Location | First-edition reading | Correction |
|---|---|---|
| Ch. 15, all six ROI boxes | Return multiples 18×, 26×, 22×, 44×, 15×, 31× | Recomputed from the stated components: 2.7×, 13×, 27×, 34×, 9.5×, 9.0× (percentages also corrected where off by a point) |
| Ch. 15, case 1 payback | 8 months | 10 months (£270K ÷ £330K/year) |
| Ch. 15, banking profile | £45M annual transaction volume | £45B (the printed figure implied ~£20/customer/year) |
| Ch. 13, optimisation box | 126× annual return; 57% cost reduction | ~10.5× first-year return; ~88% illustrative upper bound, with the tactics-overlap caveat the first edition omitted |
| Ch. 13, logging code | datetime.utcnow() (deprecated) |
datetime.now(timezone.utc) |
| Ch. 13, monitoring box | £172,800 build-up unexplained | Itemised (£144K anomaly + 3 × ~£9.6K outages) and labelled illustrative |
| Ch. 8, opening example | "£18,000 Difference" heading, £12,600 body | Reconciled and repriced (example now computes consistently) |
| Ch. 3, email costs | £0.015 / £0.006 / £0.018 per email in different places | One consistent worked example throughout |
| Ch. 6, memory ROI | 93:1 in one box, 9.9:1 in another | Single reconciled box: ~93:1 infrastructure-only, ~30:1 fully loaded (the 0.25 FTE had been charged at a full rate) |
| Intro & Ch. 2, memory ROI | 83:1 ROI, £1.67M from £20K | Arithmetic was 10× off; corrected to ~8:1 and labelled illustrative |
| Ch. 10, refund JSON | "estimated_refund_days": 3-5 (invalid JSON) |
"3-5 days" (valid string) |
| Ch. 4, model tiers | SLMs "<1B" in table, "under ~10B" in prose | Aligned on <10B / 10B–100B bands |
| Conclusion | "Five Essential Capabilities" vs "four core capabilities" | Five throughout (Memory, Tools, Reasoning, Orchestration, Safety) |
| Introduction | Composite example called a "real-world case"; gated access code promised for a public page | Labels corrected; companion page described as public and free |
| Ch. 16, quality table | "Replaceableinterchangeable" | "Replaceable, interchangeable" |
| About the author | "Wrexham Glyndŵr University" | "Wrexham University" (renamed 2023) |
Spotted one? Write to info@oldforge.tech with the page number and the issue. Confirmed corrections are listed here, with the date they were filed.
Model lineup, pricing bands, context windows and the MCP snapshot re-verified against October 2026 provider pages; sixteen first-edition errata published below; code appendix regenerated (62 snippets); new downloads section with the ROI calculator, eval harness, incident and conformity templates, an MCP server starter, and a LangGraph worked example.
AI Agents for Everyone goes live on Amazon UK and Amazon US. This companion page launches with the same April 2026 snapshot of models, pricing, context windows and the MCP ecosystem that the book is calibrated against.
Future quarterly updates (Q1 2027 and onwards) will appear above these entries. The commitment is one update per quarter; the cadence may tighten if a vendor announces something that genuinely changes the picture.
Every code and configuration snippet from the book - n8n nodes, MCP servers, tool definitions, guardrail patterns - reproduced in copy-paste form and organised by chapter. Diagrams and console-output listings are left in the book; only runnable or reference code is here.
Spans Chapters 3, 4, 5, 10, 11, 12, 13, 14 and 18 - nine chapters in all. Snippets may be reused freely with attribution to the book. Like the rest of this page, it's public - no login, no gate.
The templates, checklists and starter code the second edition points at - every file versioned, with a changelog below. Free to use; no login, no gate.
Chapter 15 template: 3-year costs, benefits, ROI and payback, plus the four-row sensitivity table. Pre-populated with the chapter's example.
Chapter 9 task-set template with scoring rubric and run log. The minimum viable quality gate before any model swap or prompt change ships.
Chapter 18 severity scale, timeline table, root-cause and comms sections. Fill it in during the response, not after.
Chapter 17 deployer checklist against the October 2026 timeline, with primary-source links. Guidance, not legal advice.
Chapter 14 minimal stdio server (tools/list + tools/call) to exercise with the Inspector before building for real.
Chapter 14 step-by-step: list, call, and break the starter server in the MCP Inspector, then graduate to a Tier-1 SDK.
Chapter 3 one-page repo rules file so coding assistants start every session with the same briefing.
Chapters 8 and 13 tracking sheet with separate cached-vs-uncached input columns, so prompt-caching savings show up explicitly.
Chapter 11 triage agent as an explicit LangGraph state machine - the code-first sibling of the book's N8N workflows.
What changed in each resource release, and when. Check here before trusting a printed version number.
About
AI Agents for Everyone sits in the gap between the books that teach engineers to build and the ones that tell executives why AI matters - written for product managers who have to make credible decisions about agentic systems, without drowning in transformer maths or marketing hype. Tone is British, direct, occasionally witty, never patronising. Pricing is in pounds.
The author, Andrei Trâmbiţaş, is the proprietor of Old Forge Technologies, a one-person workshop in Suffolk taking on bespoke software and AI-adoption work for small businesses. Two decades across DHL, Microsoft and Kaseya. MSc in Computer Science with Cyber Security. Published research on decentralised compute for large-language-model inference.