From Assistant to Agent

The concrete difference between an assistant and an agent is the size of the thing you hand it and the amount of independent action it takes in return. With an assistant, you decompose the work yourself: you ask for one output, get it, decide the next step, ask again - you are the planner and the agent between steps. With an agent, you hand over a goal - 'research the precedents for a mid-size civic library and summarise what makes them work', or 'draft the room data sheets from this schedule', or 'check this layout against these rules and list the issues' - and the agent itself breaks the goal into steps, decides what to do first, uses whatever tools it needs, adapts when a step fails, and works through to a result. The unit of delegation gets bigger, and the number of times you have to intervene gets smaller.

Three capabilities make this possible, and each is worth naming because the rest of the course builds on them. Tool use: an agent is not limited to producing text - it can take actions in the world by calling tools (searching, reading and writing files, running an analysis, operating other software), which is what lets it actually do work rather than just describe it (Module 1.2, 2.2). Planning and iteration: an agent can break a goal into steps, sequence them, and loop - try, observe the result, and adjust - rather than producing one shot (Module 1.4). And memory and context: an agent can hold the state of a multi-step job in mind, carrying what it has learned from one step into the next (Module 1.3). Put these together and you get something qualitatively new: a system you can task, not just prompt.

The practical upshot for a designer is that whole chunks of work that used to be a back-and-forth - or a tedious manual grind - can be handed over as a goal and supervised rather than performed step by step. That is a large productivity shift, and, handled well, it frees the designer's time and attention for the judgement, creativity and human relationships that are the real heart of design. Handled badly - trusting the agent blindly, delegating what should not be delegated - it is a fast route to confident, plausible, wrong work shipped under your name. Which half you get is decided by the discipline this course teaches.

Design work is unusually well suited to agentic help, because so much of it is multi-step knowledge work that surrounds a core of human judgement - and agents are good at the surrounding work while being unfit to replace the core. Think about what a real project involves beyond the creative and relational heart of it: gathering and synthesising precedents, codes, product data and site information; turning a messy brief into a structured programme; generating and comparing options against criteria; producing and coordinating drawings, schedules and specifications; checking work against rules; chasing the endless small tasks of documentation and administration. Much of this is exactly the kind of bounded, multi-step, tool-using work that agents can now take on - which is why agentic AI has a bigger claim on the design process than earlier AI did.

But the same description shows the limit, and it is the through-line of this whole course. The core of design - the judgement about what is good, the responsibility for safety and the client's interest, the creative leap, the taste, the relationships, the accountability - is human, and agents neither hold it nor can be given it. An agent can compile the precedents, but deciding which idea is right for this place and these people is yours. An agent can draft the documentation, but the professional standing behind it - the duty of care, the sign-off, the liability - is yours. So the honest picture is not 'agents design buildings'; it is 'agents do a growing share of the multi-step work around design, under a designer who directs, judges and remains responsible.' That division - agents on the surrounding work, the human on the judgement and the accountability - is the model this course teaches you to build.

This is also why the shift is an opportunity rather than a threat for designers who engage with it clearly. The parts of practice that agents can take over are, for the most part, the parts most people find least rewarding: the grind, the chasing, the reformatting. Handing those to well-supervised agents can give a designer more time for the parts that are actually the profession - the thinking, the making, the people - provided they build the skill to direct and verify agents well, which is a real and learnable craft, and the subject of everything that follows.

The single most important idea in this course, and the one that must never blur no matter how capable agents become, is that the human designer remains the author, the decision-maker and the architect of record. An agent is a tool - an unusually capable and autonomous one, but a tool - and a tool does not carry professional responsibility. When a drawing is issued, a design is approved, a building is built, it is a licensed human professional who is accountable for it: for its safety, its compliance, its fitness, its care for the client and the public. That accountability cannot be delegated to an agent, and no amount of agent autonomy changes it. The agent may do the work; you answer for it.

This has three practical consequences that run through every module. First, verification is not optional. Agents are fluent and confident, and they are sometimes confidently wrong - inventing a code clause, misreading a dimension, producing plausible nonsense (Module 8.1). Because you are responsible for the output, you must verify it, especially anything that touches safety, compliance, or a binding commitment; an unverified agent output is a liability with your name on it. Second, judgement stays with you. Deciding what is good, what is right for this project, what to accept and what to reject, is the designer's job and cannot be handed over - the agent proposes, you dispose. Third, you choose what to delegate. Some work is fine to hand to an agent and check; some - the core creative and ethical decisions, the client relationship, the final professional sign-off - should not be delegated at all, however tempting the convenience.

Holding this line is not Luddism; it is exactly what lets you use agents boldly and safely. A designer who is clear that they remain responsible can delegate freely to agents and verify what matters, capturing the productivity without the peril. A designer who forgets it - who lets the agent's fluent confidence stand in for their own judgement and duty - is the one who ships the dangerous error and discovers, too late, that 'the AI did it' is not a defence. Use agents as powerful instruments of your practice; stay, always, the professional who directs them and answers for the result.

This course builds agentic design as a practical, professional skill. You will start with why agents change design - the shift, what an agent is, levels of autonomy, the landscape (Module 0); then understanding agents - how LLMs became agents, tools, memory, planning (Module 1); the agentic toolkit - instructing agents, giving them tools, grounding and retrieval, orchestration (Module 2); agents in the design process - research, ideation, the brief, site (Module 3); documentation and technical work - drawings, specs, schedules, compliance (Module 4); agents and the model - BIM, generative modelling, analysis, visualization (Module 5); multi-agent and autonomous workflows - agent teams, pipelines, human-in-the-loop, how far autonomy should go (Module 6); building your own agents - no-code to custom, MCP, reliability (Module 7); judgment, ethics and control - verification, the architect of record, IP and data, bias and liability (Module 8); agents in practice and the studio - adoption, workflow, the studio system, cost and ROI (Module 9); and the autonomous future and your role (Module 10).

Two honest boundaries run through it. First, agentic AI moves extraordinarily fast, and any specific tool, model or capability named here is illustrative and dated - the durable value is in the principles (what an agent is, how to direct and verify it, where to keep a human), which outlast any particular product. Treat tool names as examples to verify against the current state, as of 2026, not as recommendations frozen in time. Second, and more importantly, this course never defers professional judgment, responsibility or the duty of care to a tool. It is emphatic that you remain the author and the architect of record; that agent output must be verified, especially anything touching safety, code or a binding commitment; and that the core creative and ethical decisions of design are yours to make and answer for. Agents are instruments of your practice, not replacements for your judgement.

Studio Matrx is free and not-for-profit, and this course is written to be clear-eyed rather than either breathless or fearful - to help architects, interior designers and students genuinely understand this fast-moving shift and put it to work in a way that makes their practice better and keeps it professional. Learn to direct agents well and verify them rigorously, and you gain a powerful new capability while remaining, exactly as before, the human who designs, decides and is responsible.