Don't Blame the Model: Seven Ways to Inject Context and Control into Agents

bibryam · x · 2026-08-16

When an agent underperforms, we tend to blame the model, then inflate the system prompt — and when that becomes unwieldy, we bolt on another agent. The author argues both skip a more useful design question: what information or control is the agent missing, and where should it enter the run?

The post maps seven common ways to add context or control: user prompts, system prompts, scoped instructions, skills, tools, steering, and delegation. They differ along four dimensions: when they enter, how far they reach, how long they persist, and whether the work stays in the current context.

Examples use Claude Code, but the patterns apply across agent frameworks. Diagrams track two clocks — when the runtime resolves each source, and how model input evolves across successive calls, showing which contexts form a stable prefix. The takeaway: picking the right injection point beats enlarging prompts or stacking agents.

Original post →

Don't Blame the Model: Seven Ways to Inject Context and Control into Agents

More from coding & agent