Make the delegation inspectable
Delegating a job to an agent is a design decision about authority, outputs, cost, state, review and parallel work. These two companions let you check the series' worked examples line by line instead of taking the conclusions on trust.
Free companions · v0.1 · October 22, 2026
Illustrative, not measured.
The dossier is entirely fictional: its sources, claims and decision are invented to show the shape of a usable brief. The cost model uses hypothetical inputs with reproducible arithmetic. Neither reports an executed benchmark, a vendor price or an achieved saving.
Six fully written source cards, a claim register and the decision they support, including the ambiguity that is resolved and the one that is not. Pairs with the research-assistant article and the continuation example in the memory article.
Inputs and derived values for the cost comparison: human review, correction, setup, maintenance, volume sensitivity and break-even. Change an input and the article's conclusion can change with it; that is the point.
Read the series in order
- How much autonomy does this job need?Allocate decisions and make the handoff inspectable.
- Build a research assistant whose work you can useProduce a decision brief with traceable facts, uncertainty and reversal conditions.
- The cheapest agent can be the most expensive workflowCompare accepted outcomes including human work, setup and maintenance.
- What does an agent need to remember?Preserve the evidence and authority needed to resume correctly.
- Your coding agent's reviewer needs evidenceDerive checks from intended behavior and distinguish findings from execution.
- When more agents are worth itFind independent work and account for coordination.
What the series does and does not claim
The articles cite published studies and vendor engineering reports as specific examples, not as proof that the proposed designs work. Sample selection and denominators are preserved where a study is quoted. No article claims an executed benchmark of its own.
For executed observations, start with the filesystem lab and the AgentDojo reanalysis.