Docs/Guides/Economic Policy

Start at 1 Credit: A Practical Economic Policy for Agents

Don't guess tool prices on day one. In Observe mode, price every route and tool at 1 credit first, measure real usage, refine policy from evidence, then move to Control with confidence.

The problem with guessing too early

Most teams start economic policy backwards. They try to assign “realistic” prices to routes and tools before they know what agents actually use. That creates noise immediately. If one tool costs 10 credits and another costs 1, you can't tell whether usage patterns reflect real demand or just your arbitrary pricing.

SatGate works better when Observe mode is used as measurement first, policy refinement second, and enforcement last.

The recommended rollout

  1. Set every route and tool to 1 credit in Observe mode.
  2. Watch raw usage patterns across agents, tools, and routes.
  3. Compare observed usage to provider invoices from OpenAI, Anthropic, image APIs, search APIs, and internal services.
  4. Identify true cost drivers, not just high-frequency actions.
  5. Reweight expensive or risky actions once you have evidence.
  6. Refine behavior while still in Observe by tuning prompts, routes, batching, caching, and workflow design before Control mode starts enforcing budgets.
  7. Move into Control mode with budgets that steer behavior intentionally.

Rule of thumb: In Observe, optimize for signal, then refine policy and behavior. In Control, enforce what worked.

Why 1 credit works

  • It creates a level playing field. Every route and tool starts with the same weight.
  • It reveals real behavior. You see what agents choose when pricing does not distort the result.
  • It gives you a baseline. You can map raw usage against real-world invoice data.
  • It avoids premature policy mistakes. You don't punish the wrong actions based on guesswork.

How to use the data

Once you have a few days of Observe data, don't jump straight to enforcement. First use the data to refine policy while the system is still non-blocking. Look for two things:

  • High-frequency actions, which show what agents naturally rely on
  • High-cost actions, which show what is actually driving your bill

Those two are not always the same. A tool may be used rarely but dominate spend. Another may be used constantly but be cheap. SatGate lets you separate those signals, then encode the difference into policy.

That is the missing middle step: use provider usage and cost analysis against SatGate data to start changing agent behavior during Observe mode. Reprice tools, tune prompts, route work differently, and reduce waste before Control turns those decisions into enforcement.

From Observe to refined policy to Control

After the first Observe pass, make expensive tools cost more credits and keep cheap tools light, but stay in Observe long enough to see how behavior changes. Then assign budgets to teams, agents, or delegated tokens in Control.

That changes behavior naturally:

  • Humans avoid burning budget on low-value actions
  • Agents learn that some tools are more expensive than others
  • Teams start optimizing prompt design, batching, caching, and workflow design

This is the real value of economic policy. You are not just blocking spend. You are shaping better behavior before the invoice shows up, then enforcing a policy already informed by real usage.

Example

Week 1: Observe
- web_search = 1 credit
- file_search = 1 credit
- code_exec = 1 credit
- image_generation = 1 credit

Result:
- image_generation is only 4% of calls
- but it drives 46% of provider cost
- web_search is 38% of calls
- but only 8% of provider cost

Week 2: Observe, refined from data
- web_search = 1 credit
- file_search = 1 credit
- code_exec = 2 credits
- image_generation = 12 credits
- prompts and workflows are tuned to reduce waste
- agents begin adapting before enforcement

Week 3: Control
- budgets now deplete in proportion to actual economic impact
- enforcement lands on a policy already tested in Observe.

SatGate recommendation

If you do not know what a route or tool should cost yet, set it to 1 credit in Observe mode. Let usage teach you. Then turn that evidence into refined policy before moving to Control.

Start simple. Measure first. Refine second. Enforce last.

Related