Cost and latency budgets for Agents that run every day

A one-off experiment can cost whatever it costs. A scheduled Agent is a subscription you wrote yourself. This lesson is about knowing the price before the bill arrives.

Estimate before you build

For each model step, three numbers:

Multiply, then multiply by 30. If the monthly number makes you pause, redesign now; it is ten times cheaper than redesigning after launch.

Measure after

Log, for every run: tokens in, tokens out, model, duration, and whether retries happened. Most workflow tools expose these in the step output; write them to a sheet. Look at the 95th percentile, not the average. The heavy days are the ones that set the bill.

The three cheapest fixes

1. Send less. Strip HTML, signatures and quoted replies before the model step. Cap excerpts. Half the tokens most workflows send are boilerplate.

2. Route by difficulty. Use a small, cheap model for the "is this relevant at all?" filter, and the capable model only for the items that pass. Two steps often cost less than one.

3. Cache what does not change. The same article summarized twice is money burned. Key your cache by URL or content hash and skip the call when you have seen it.

Latency

Daily digests do not care about seconds. Interactive tools do. Two rules:

• Set a timeout on every model step (30 seconds is generous for a summary) and a fallback ("summary unavailable, here is the raw item").
• Run independent model calls in parallel when the tool allows it. Summarizing ten items one after another is ten times slower than it needs to be.

A budget line for the project page

Every advanced project here states its cost per run and the assumptions behind it. Do the same for yours, in one line, at the top of the workflow description. It is the fastest way to keep a project honest.