Drills: Agents.
Five questions at a time from the whole volume, until every agent design choice has its trade-off attached.
This plate draws rounds of five from the questions and predictions of the eight plates before it, with the reasoning and source plate after each answer.
The volume’s thread is that an agent is a model in a loop with tools: it extends what the model can do, and every turn adds context, cost, a chance to fail and a chance to be misled.
Below: a round of five. Answer each, read the explanation, and press “Next round” for another five.
- Explain tool calls and the agent loop.
- Choose retrieval and context strategies.
- Reason about reliability over long tasks.
- Lay out defences for acting agents.
What to carry forward
A tool call is text a harness executes (plate XXV.1); an agent acts on observations in a loop (plate XXV.2); retrieval brings evidence into the context (plate XXV.3); and the context must be managed because it grows and is re-read every step (plate XXV.4).
Errors compound as pⁿ unless repaired (plate XXV.5); screens make every application reachable at a cost in precision (plate XXV.6); many agents buy breadth with tokens (plate XXV.7); and acting agents need layered defences (plate XXV.8). Vol. XXVI turns to how agents are trained.
Terms
- Agent
- A model in an act–observe loop.
- Harness
- The software that runs tools.
- BM25
- Lexical ranking.
- pass^k
- Success on all of k tries.