TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 18—21 · FACTUAL · DURABLE
The blast radius of an error
One misread level name becomes fifteen wrong sheets, because nobody reads the middle.
The idea
A chatbot's error is one wrong answer in a box. You read it, you catch it or you do not, and it goes no further on its own. An agent's error is different in reach, not in kind. The agent takes its wrong answer as the start of its next step, and the next, and nobody reads the middle. One misread level name becomes fifteen wrong sheets. Blast radius is the measure: how much a single wrong step can touch before a person looks. It grows with permissions, autonomy and the time before anyone checks.
Why it matters
The size of a mistake is not the size of the error. It is the size of everything the error could reach before it was seen.
See it in the studio
An agent is asked to update every sheet that references the ground-floor level after a change. It reads the new level from the wrong drawing. Every referencing sheet is now consistently, tidily wrong — and consistency is what a checker looks for.
Watch for this
Judging an agent by the quality of its answers. Judge it by the reach of its actions. A brilliant agent with a large blast radius is more dangerous than a mediocre one with a small one.
Try it
Take one automated workflow you use or plan. Draw its steps as a chain. Mark the first step where a wrong output is consumed by the next. Count the steps from there to the first place a human looks. That number is your blast radius. Shrink it.
Prove it
Explain why the same error rate means a different risk in a chatbot and in an agent, using one chain of steps from your own work.
How it works
Errors compound along a chain. If each step is right ninety-five times in a hundred, a chain of ten unchecked steps comes out right only about sixty times in a hundred. That is arithmetic, not opinion, and it is why the people building these systems recommend short workflows with checkpoints over long autonomous runs. Two things shrink the radius: shorten the chain before a human reads, and make a wrong step stop the chain rather than feed it. Neither needs a smarter agent. Both need the workflow to be designed.
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.