SKIP TO CONTENT
AI AGENTS6 minAgents in Production / Ep. 7 · Stage 7-3

AI Agent Fleet Costs: What to Cut First

Someone hands you an agent fleet’s bill and asks you to halve it. Would you start with the model?

XIN

IN 30 SECONDS

Cutting an agent fleet’s cost is mostly design work: how much each agent may do alone, how often it hands cases to people, and when it retires. The model comes last, because in this post’s model tokens are a small line.

KEY TAKEAWAYS · 3

  • In this illustrative model, review and upkeep are the big lines; they track an agent’s autonomy more than its tokens.
  • The Integra agents and the 11-agent fund platform are two separate systems. Their counts are never added together.
  • The cheapest agent is the one you retire. Plan for retirement from the start; it is a design feature, not an admission of failure.
On this page

Autonomy can drive an agent’s cost

IN SHORT What an agent may do on its own says more about its cost than what it’s for.

WORDS TO KNOW

Autonomy tier
How much an agent may do on its own, from only reading data to changing live systems.
Sandbox
An isolated environment set apart from live systems.
Escalation
The agent handing a case it is not sure about to a person, with what it found.

About 46 agents ran internal work at Integra, where I was CTO until June 2026. The fund platform, with 11 agents at first delivery and 25 modules at handover, is a separate system with its own case study, and I never add the two counts.

I keep the function-by-function split private. Instead, I sort agents by Autonomy tier How much an agent may do on its own, from only reading data to changing live systems. See the glossary: how much each may do without a person. These tiers are a teaching model, not a published inventory of the Integra fleet.

THE FLEET, BY AUTONOMY TIER
Layered stack diagram: 4 layers, top to bottomLayered stack diagram: 4 layers, top to bottom. Layer 1, TIER 0 · READS ONLY: Alert triage, Log digests. Layer 2, TIER 1 · DRAFTS, A HUMAN SENDS: Status reports, Ticket routing. Layer 3, TIER 2 · ACTS IN A SANDBOX: Scheduled data pulls, Undoable automation. Layer 4, TIER 3 · TOUCHES PRODUCTION: Human gate on every action, Never acts alone.TIER 0 · READS ONLYAlert triageLog digestsTIER 1 · DRAFTS, A HUMAN SENDSStatus reportsTicket routingTIER 2 · ACTS IN A SANDBOXScheduled data pullsUndoable automationTIER 3 · TOUCHES PRODUCTIONHuman gate on every actionNever acts alone
In this teaching model, the same summary task costs little at tier 0 and can cost far more at tier 3, where a person approves every action.

IN PLAIN WORDS

In this teaching model, tier 2 runs undoable automation in a Sandbox An isolated environment set apart from live systems. See the glossary. An action that needs human review can cost more to run than the same model used read-only. Measure how often review happens and how long it takes. Don’t assume a fixed multiplier.

I’d set the tier at build time and budget for raising it later. Where that boundary sits is the subject of least privilege for agents that move money.

Coming up next: So where would I start cutting a fleet’s bill?

Cut the agents nobody reads first

IN SHORT Delete what nobody reads first, then fix the design; the model comes last.

If someone asked me tomorrow to halve a fleet’s bill, I wouldn’t start with the models. I’d open the list of agents and work in this order.

STEP BY STEP

Decide: An agent costs more than it should. What do you do?Decision tree: An agent costs more than it should. What do you do? If nobody has read its output in 30 days, then Delete it. If the review line is the biggest, then Fix the escalation rate, or lower the tier. If the engineering line is the biggest, then Fix the shared part, not the agent. If the token line really is the biggest, then Use a cheaper model, cache, batch.DECIDEAn agent costs more than itshould. What do you do?If nobody has read itsoutput in 30 daysDelete itIf the review line is thebiggestFix the escalationrate, or lower thetierIf the engineering line isthe biggestFix the shared part,not the agentIf the token line really isthe biggestUse a cheaper model,cache, batch

Step 1. Nobody reads it. Delete it. The cheapest agent is the one that no longer exists, and a fleet with no retirement rule only grows.

Pick a step to see it move
  1. Nobody reads it Delete it. The cheapest agent is the one that no longer exists, and a fleet with no retirement rule only grows.
  2. Review is the biggest line The “I’m not sure” bar is too cautious, or the tier doesn’t fit. Repeats are bugs to fix. Lowering the tier can cut risk but add review: measure first.
  3. Engineering is the biggest line Recurring upkeep on one agent often hides a platform problem. Fixing the shared part once may help every agent on it.
  4. Tokens really are the biggest line Rare, and the easiest case: a cheaper model, caching (reusing results) or batching. A cheaper model must still pass the same quality tests.
The order I’d use. Three of the four branches are design work; measure the savings in each, since they differ between fleets.

Two of these fixes, the Escalation The agent handing a case it is not sure about to a person, with what it found. See the glossary rate and the tier, are design choices.

An agent fleet is mostly a staffing decision dressed up as infrastructure.

You’re buying the duty to keep 46 things working, not tokens. Left unmeasured, review and upkeep can turn someone into a full-time reviewer nobody chose to hire.

ITEM GET

Retirement comes first

Give every agent a retirement rule from day one, such as “nobody has read its output in 30 days”.

FAQ

In stage 1’s illustrative model, one internal-ops agent costs about $26 a month in tokens and $521 in all. Those are modelled figures, not an invoice. What matters is the split: tokens about 5%, review about 58%, and yours will differ.

No. About 46 agents ran internal work at Integra. The confidential fund platform, through Ilayer, had 11 agents at first delivery and 25 modules at handover; the work overlapped in 2026, but these are two systems at two stages.

Retire the workflows nobody uses, then look at your biggest measured cost. Less autonomy can lower risk but add human review. Weigh that against a model or integration change before you claim a saving.

STAGE CLEAR

What you now know

  • Why, in this teaching model, what an agent may do alone says more about its cost than what it’s for.
  • The order I’d cut a fleet’s bill in, and why the model comes last.

Check yourself

Tap or click a question to see the answer.

Same model, same task. One agent drafts replies a person sends (tier 1); the other changes live records, a person approving every action (tier 3). Which likely costs more to run?

Probably tier 3: every action waits for approval, and review is the expensive line. Only measured review time settles it, with no fixed multiplier.

Next episode:What should you record when an AI agent acts, so an auditor can trace one output later?6 min
Adam Boudjemaa
Adam Boudjemaa

Former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft. Building production AI and regulated Web3 systems.

Have a project to ship or a role to fill?

Discuss your project or role
LinkedInTwitterGitHub
Follow the engineering notes

Get new articles on production AI and smart contracts by email.