Agent FinOps for Design Orgs

Design orgs discover agent cost in the wrong order. First comes delight at faster drafts. Then parallel critique agents, full-page screenshots stuffed into context, and sprawling design-system files loaded on every turn. Then finance asks why the team seat invoice and the API overage both jumped in the same month.

Agent FinOps for design is not about denying people tools. It is about knowing which loops are expensive, which checks must stay fail-closed anyway, and which habits burn money without improving decisions. Visual work is heavier than code-only work: images, repeated captures, and long DESIGN.md context change the arithmetic.

This field note pairs with plan-selection guidance elsewhere on the site. Here the focus is operational: measure, budget lanes, and put cheap gates before expensive judges.

Five buckets explain most invoices. Context load is the silent one: huge instruction files and entire component trees pulled into every session. Generation is visible and usually not the villain. Capture is sneaky when screenshots enter the model context as base64 instead of living on disk. Critique multiplies cost when three agents each re-read the same images. Retries after vague prompts multiply everything.

A product design team ran a "quality bar" experiment: every marketing landing PR spawned three critique agents (visual, a11y, brand) with full-page desktop and mobile screenshots inlined into each prompt. In two weeks they caught real issues — and spent more on critique than on generation. One recurring blocker, a clipped mobile CTA, was something a deterministic visibility assert would have caught for nearly nothing.

They changed the order. Capture wrote files to disk. A Playwright check failed on CTA visibility and axe serious findings. Only passing candidates reached a single judge agent with cropped regions, not three full-page triplicates. Monthly spend dropped; blocker escape rate did not rise. The FinOps win was sequencing, not cheaper models.

Cost control should feel like craft, not rationing theater.

You do not need a full FinOps platform on day one. You need a weekly fifteen-minute skim and a written budget for expensive lanes.

Owner: [design eng manager]

Inputs:
- Seat plan usage / credit burn (vendor dashboards)
- API/BYOK invoice if any
- Count of eval:visual and critique runs this week
- Top 3 longest agent sessions (ask the team)

Questions:
1. Did critique spend exceed generation spend?
2. Which loops inlined images instead of using disk captures?
3. Which failures should move to deterministic gates?
4. Any BYOK key without an owner?

Actions (max 3):
- [ ] gate change
- [ ] prompt/context trim
- [ ] budget or plan change

Log path: ops/agent-finops/YYYY-MM-DD.md

Do not save money by deleting accessibility scans or by skipping mobile captures. Save money by refusing duplicate judges, oversized context, and conversational retries that should have been a rewritten brief. Prefer subscription tiers that match daily design-to-code intensity over surprise BYOK bills nobody watches.

Publish a simple policy: fail-closed structural evals are mandatory; multi-agent critique is opt-in above a PR size threshold; screenshot-to-model requires crops. Agents should read that policy from AGENTS.md so the cheapest correct path is the default path.

• Mandatory: capture on disk, axe on changed interactive routes, named spend owner.
• Metered: LLM visual judge, multi-agent critique fleets, huge context dumps.
• Forbidden as default: base64 full-page images into three parallel critics.
• Revisit budgets monthly; tooling prices move faster than org habits.