AgentForge

Cost Control

Practical strategies for controlling agent costs — budgets, model selection, and monitoring.

Strategy 1: Set Budget Limits

Always set explicit limits:

budget: {
  maxCostUsd: 1.00,      // Hard cap
  maxTokens: 100_000,    // Token cap
  warnThreshold: 0.7,    // Warn early
},

Strategy 2: Choose the Right Model

TaskRecommended ModelRelative Cost
Simple queriesClaude Haiku / GPT-3.5$
Standard tasksClaude Sonnet / GPT-4o$$
Complex reasoningClaude Opus / GPT-4$$$$
// Use Haiku for simple tasks
const cheapLLM = createLLM({ provider: 'anthropic', model: 'claude-haiku-4-5-20251001' });

// Use Sonnet for standard tasks
const standardLLM = createLLM({ provider: 'anthropic', model: 'claude-sonnet-4-20250514' });

Strategy 3: Limit Steps

const agent = defineAgent({
  // ...
  maxSteps: 5,  // Prevent runaway loops
});

Strategy 4: Monitor with Hooks

hooks: {
  onStep: (step) => {
    metrics.recordTokenUsage(step.tokens.input + step.tokens.output);
  },
  onComplete: (result) => {
    metrics.recordRunCost(result.trace.summary.estimatedCostUsd);
  },
},

Strategy 5: Escalation Thresholds

policy: {
  costEscalationThreshold: 0.50,  // Human review at $0.50
},

Strategy 6: Multi-Agent Budget

const coordinator = new Coordinator({
  agents: { researcher, writer },
  budget: {
    maxTotalCostUsd: 3.00,
    maxCostPerAgent: 1.50,
  },
});

Next Steps

Production Deployment

Deploy AgentForge agents to production — state, monitoring, scaling, and best practices.

Human-in-the-Loop

Implement approval workflows — escalation, pausing, and async approval patterns.

On this page

Strategy 1: Set Budget LimitsStrategy 2: Choose the Right ModelStrategy 3: Limit StepsStrategy 4: Monitor with HooksStrategy 5: Escalation ThresholdsStrategy 6: Multi-Agent BudgetNext Steps