Why agent token costs are the new cloud bill

A few years ago, "the cloud bill" became a boardroom topic. Compute scaled with your product, nobody quite knew where the money went, and a whole category of tooling — cost dashboards, rightsizing, reserved instances — sprang up to tame it.

Coding agents are on exactly the same trajectory, only faster. The more your engineers lean on Claude Code, Codex and opencode, the more your token bill grows. And unlike a runaway EC2 instance, an agent's spend is supposed to grow — it scales with productivity. The problem isn't that agents cost money. It's that almost nobody can see or shape that cost.

The anatomy of an agent invoice

Open up where the money actually goes and you'll find three culprits, every time:

1. Prompts are enormous — and repetitive

Every agent turn re-sends context: the file tree, relevant files, tool output, and the entire prior conversation. A single "fix this test" request can carry tens of thousands of input tokens, most of which the model has already seen on previous turns. You pay for all of them, every time.

2. Everything runs on the flagship model

Agents default to the most capable model available — because occasionally they need it. But most requests aren't "design a distributed scheduler." They're "rename this variable," "add a docstring," "fix the import." Paying flagship rates (think $3 / $15 per million tokens) for work an economy model handles at $0.20 / $0.60 is pure waste.

3. There's no visibility until the invoice lands

Who on the team is spending what? Which repo? Which agent? For most orgs the answer arrives once a month, in the form of a single large number with no breakdown.

The more valuable agents become, the more they deserve a real infrastructure layer — one that makes them cheaper to run without changing how developers work.

The three levers

The good news: each culprit has a direct counter-measure, and they compound.

• Compress the prompt. Strip repetition and distill verbose tool output before it reaches the model — roughly halving input tokens with no loss of meaning.
• Route each request. Score it for complexity and send the simple majority to an economy model, reserving the flagship for genuinely hard work.
• Govern the spend. Per-user budgets, virtual keys and a live activity feed so cost is visible and capped in real time — not reconciled after the fact.

Do all three at the single point where every request already flows — between the agent and the model — and you cut the bill 60–90% without asking a single developer to change their workflow.

That's what we built

Cohesor is the gateway that applies all three levers automatically. If your agent spend is starting to look like a cloud bill, that's not a reason to slow down on agents — it's a reason to put infrastructure underneath them.