02. Prompt Length, Context Growth, and Cache Cost
A trace-based analysis of how multi-turn context grows and what prefix caching, trimming, summaries, and server-side state references actually change.
00. Reading the Agent Runtime from Its Traces
A reproducible route from captured model requests to tool events, permission decisions, recovery behavior, and Agent evaluation.
13. Multi-Agent Value Is Not Multiple Models Running
Evaluating delegation through task packets, context isolation, reduced permissions, result verification, and total coordination cost.
14. Recording a Replayable Agent Trace
Joining model, tool, permission, and state events with trace, turn, call, and sequence identifiers to reconstruct causality.
05. What Happens After the Model Calls a Tool
Following one tool request through parsing, schema validation, authorization, execution, event recording, and result delivery.
03. Where Prompt Injection Enters an Agent
Placing adversarial instructions in user input, project files, tool results, retrieval, and memory to separate model interpretation from enforceable permissions…
11. Which Layer Should Handle an Agent Error?
Separating model, authentication, quota, protocol, validation, permission, tool, and result errors into different recovery policies.
08. Why an Agent Needs a State Machine
Representing model, authorization, and tool execution as explicit states so interruption recovery does not depend on conversation text.