Prompting And Planning

Smart agents shape model behavior through a generated system prompt. That prompt is where autonomous planning and execution discipline are taught.

buildSystemPrompt(...)

import { buildSystemPrompt } from "@cognipeer/agent-sdk";

const prompt = buildSystemPrompt(
  "Keep answers short and cite sources.",
  "todo",
  "ResearchHelper",
);

What the prompt adds

• agent identity
• general execution rules
• planning rules when planning is enabled
• optional extra instructions from systemPrompt

Planning instructions matter most for autonomous agents

When planning is enabled, the prompt teaches the model to:

• create a plan only when the task is genuinely multi-step or explicitly requested
• skip planning for direct Q&A or one-step lookups
• use write only for initial creation or full rewrite
• prefer update for status, evidence, and progress changes
• include expectedVersion when possible
• recover from version mismatch with read plus retry instead of destructive rewrite

These are the behaviors that make smart planning useful for autonomous agents rather than just decorative task narration.

Adding extra instructions

const agent = createSmartAgent({
  model,
  tools,
  systemPrompt: "Prefer terse answers and mention tradeoffs explicitly.",
});

Use this for domain-specific constraints, tone, output expectations, or operational policy.

Full prompt escape hatch

If you provide a leading system message in invoke(...), the smart runtime will not prepend another one. That is the right escape hatch when you intentionally want full system-prompt control.