AI Agent Architecture Cheat Sheet
Every moving part drawn out: reasoning, memory, retrieval, tools, guardrails, and the point where a person signs off.
ImageFirm advises governments, global enterprises and frontier founders on ethical AI — and builds the models, agents and creative work that prove it can be done right. Every sheet here is that work, in the open. Start anywhere.
ImageFirm is an ethical-AI advisory and applied-AI studio. It advises governments, global enterprises and frontier founders, builds production models and agents inside companies it owns and runs, and publishes the working documents — seventy-one sheets across seven disciplines — in the open, under five principles it does not negotiate.
This library is ImageFirm's first public asset — the start of the journey. Every sheet is a working document from things we actually build and run: agents, rollouts, intelligence work, field projects. It is published in the open, with no paywall, because that is how we believe AI should be built — human-first, transparent, accountable. The index will keep growing. Start anywhere. If it helps you build, we are easy to find.
— Uzay Kadak, Founder
The three sheets to read first — one per funnel stage.
Every moving part drawn out: reasoning, memory, retrieval, tools, guardrails, and the point where a person signs off.
Seven systems — strategy, value, organisation, people, governance, engineering, data — sequenced into one plan.
Search the full list, compare tools side by side, or let it assemble a five-tool stack for you.
Every sheet, filed by discipline. Open a drawer.
How an agent is put together, and what has to be true before it runs unattended.
One sheet for the person using it, one for the person approving the budget.
Sequencing, governance and the engineering discipline that keeps a rollout from stalling.
Reading signals, judging exposure, and staying composed while you do it.
Searchable lists, kept current, for when you need the tool rather than the theory.
Where the frameworks meet a real operation with real consequences.
Onboarding material for people joining a team that already works this way.
Five principles we don’t negotiate — on every sheet and in every engagement.
The most recent additions to the library.
A-18Agentic AI Engineering2026-09-11
L-05How Much Memory Do You Need to Run a Local LLM? — ImageFirm Research2026-09-08
A-16The Agent Go-Live Gate — 24 questions before an AI agent runs unattended2026-09-06
A-17Enterprise AI Agent Readiness & ROI Framework 20262026-09-06
Short answers. The sheets carry the long ones.
A working library of seventy-one one-page documents across seven disciplines — agent engineering, Claude Code, enterprise strategy, intelligence, directories, field projects and learning — published in the open by ImageFirm. Every sheet stands on its own. This page is Sheet 00, the master index.
Two readers. Leaders deciding how to govern, sequence or buy AI — start with the enterprise and field sheets. Operators building agents and running Claude Code — start with agent engineering and the directories.
Three disciplines under one standard: ethical-AI advisory for governments and enterprises; applied AI and agentic systems built to production; creative and content work for organisations at the frontier. Advisory work is confidential and stays that way.
Each sheet is drafted with AI assistance and verified by a person before it is published. It carries a sheet number, a discipline and a revision stamp. The index and the discipline pages are generated from one manifest, so counts and links stay true.
Read, cite and link freely. Write to us before republishing them or building a product on them.
In batches. The revision stamp in the footer and the latest-additions strip above show when.
Three ways in: the Field Notes letter, a thirty-minute executive briefing on one sheet applied to your situation, or a private session for a selective engagement. A person reads every enquiry and replies within two business days. No lead funnels, no automated scoring.
No pop-ups, no retargeting, no automated scoring. Traffic is measured in aggregate only.
No lead funnels, no automated scoring, no retargeting. Your message is read as a letter, not a data point.