context-fundamentals

Organize context components—system prompts, tool definitions, retrieved documents, message history, and tool outputs—into a structured hierarchy. Clear separation and logical ordering help agents interpret instructions consistently and reduces behavioral drift caused by context conflicts or ambiguity.

Token budgeting allocates fixed space across context components using progressive disclosure: prioritize high-value system prompts and tool definitions, compress message history, and defer non-critical documents. This maintains agent performance under token limits while preserving decision quality.

Agent behavior shifts when context components—especially system prompts, tool outputs, and message history—interact unexpectedly or compete for attention. Debugging requires mapping all five context components, observing their relative size and order, then adjusting each to isolate the root cause.

Optimize token efficiency by right-sizing each context component: compress verbose tool definitions, trim message history to recent turns, summarize retrieved documents, and prune redundant system instructions. Progressive disclosure ensures the agent receives only necessary context per task.

System prompts guide overall agent reasoning and constraints; tool definitions specify available actions and their parameters. Both occupy token budget but serve distinct roles—system prompts shape strategy, while tool definitions enable execution. Balancing their size and clarity prevents conflicts.

Yes. Systematically isolate failure causes by mapping context components, then incrementally adjust each—removing tool outputs, shortening history, simplifying prompts—and observe behavior changes. This reveals whether failures stem from prompt ambiguity, tool confusion, or context overload.