Goal-Based Agents

Goal-based agents do not just react. They choose actions that help reach a target state.

This is the "planner" mindset: define a goal, evaluate possible next steps, and execute the best path.

How it works (architecture)

1. Define goal and constraints.
2. Generate plan candidates.
3. Select next action toward goal.
4. Execute and observe outcome.
5. Re-plan until goal is reached or aborted.

Typical examples of goals

• "Create a market brief in 30 minutes with sources."
• "Automate invoice matching with less than 1% error."
• "Book meetings with qualified leads this week."

Best use cases

• Task completion workflows
• Planning across multiple steps
• Structured assistant behavior
• Research and synthesis pipelines
• Agentic coding and refactor workflows

Trade-offs

• Needs clear goal definition
• Planning cost can grow quickly
• Poor goals create confidently wrong behavior

Real-world company and service examples

• OpenAI - goal-driven assistants built with the Responses API + tools (web search, files, code execution, external APIs).
  Approximate API range: $0.20/$1.25 to $2.50/$15 per 1M input/output tokens for many standard model tiers.
• Anthropic Claude - planning workflows with tool use for research, drafting, and operations support.
  Approximate API range: often $1/$5 to $3/$15 per 1M input/output tokens for common tiers.
• Microsoft Copilot / GitHub Copilot style task completion flows (goal = "produce result, not just answer").
  Approximate pricing: commonly $10-$39 per user/month depending on product tier.

Practical pricing references (consumer + API)

• ChatGPT Plus: about $20/month (web app subscription; API billed separately).
• Claude Pro: about $20/month monthly, or around $17/month annual equivalent in some regions.
• API cost can be lower than subscriptions for light usage, but higher for heavy automated workloads.

How to make goal-based agents reliable

• Write explicit success criteria ("done means X").
• Enforce budget caps (tokens, time, tool calls).
• Add checkpoints after every major step.
• Require citations/evidence for factual tasks.
• Add human approval for high-risk actions.

When to avoid

Avoid pure goal-based setups when: - You need strict optimization under multiple conflicting objectives (use utility-based). - Environment changes continuously and model must adapt from feedback (add learning components).