TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 19—21 · FACTUAL · EVOLVING
Agents, goals, tools, memory
An agent acts over several steps, so decide in advance where the human gate sits.
The idea
An agentic system is a model placed inside a loop with some authority to choose its next action from the result of the last one. Give it a goal and it works towards it over several steps. Planning, tool calls, stored state and self-checks are parts it may have; none alone makes it an agent, and a fixed sequence of model calls laid down in code — a workflow — is not one either. An agent acts over time, which buys reach and costs reliability: errors add up. So checking is built in: runtime gates where a human approves.
Why it matters
When a machine acts over several steps, responsibility is no longer a single moment. You have to decide in advance where the human gate sits, because afterwards is too late.
See it in the studio
"Prepare the area statement for the municipal submission." The agent opens the model file, pulls areas, formats the table, checks totals against the byelaw limit, drafts the cover note — and pauses before filing. That pause is the design, and it is a permission gate in the runtime, not a sentence in the goal. "Stop before filing" written in the brief asks; the runtime's gate enforces. The closer a step comes to a signature, a payment or something that cannot be undone, the stronger the gate should be.
Watch for this
An agent that reports success in fluent prose. The report is generated like everything else. Read the artefact, not the summary.
Try it
Take one studio task and write out its loop by hand: goal, first tool call, what state must be kept, what self-check runs, where a human must approve. Then decide which step you would never automate, and say why.
Prove it
Explain the difference between asking a model a question and handing an agent a goal, name the parts of the loop for one studio task — goal, tool call, memory, self-check — and say which of them that task actually needs, and where the gate sits.
How it works
The loop is orchestration code around the model. It supplies tools (TOOL USE on this strand), stores intermediate state, and feeds results back as context. Practitioner guidance distinguishes workflows, where code fixes the path, from agents, where the model directs its own process, and favours the simplest form that does the job; it treats permissions as the real design decision, and they live in the runtime, not in the prompt. EVOLVING: agent frameworks and their reliability are moving fast. What counts as "built-in checking" this year may be standard plumbing next year.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.