Claude Code Field Notes · Notes · 5 of 5

Cost control for coding agents: budgeting tokens on fixed-fee South Florida web projects

Note 5 of 5 in Notes from West Palm Beach, published November 15, 2026.

A fixed-fee website project has two numbers: the one the client signs and the one the agency absorbs. When coding agents do part of the work, token spend joins the second number, and nobody notices it until a small scope change turns into an afternoon of an agent retrying the same failing test.

A lot of small-business web work in West Palm Beach, Jupiter and Boca Raton is sold at a fixed fee: a brochure site, a booking flow, a migration off an old page builder. This note is how Jason Colapietro budgets agent usage on that kind of work so the margin survives the project. It is a method, not a price list. Model prices change, and the arithmetic below works with whatever your plan charges.

Where the tokens actually go

Writing code is rarely the expensive part. On Suede's own projects, spend clusters in four places:

Three of the four are about context, not intelligence. That is good news, because context is the part you control.

Budget per task, not per project

A project-level budget tells you that you overspent after it happened. A task-level budget stops the overspend while it is happening. Jason Colapietro breaks a fixed-fee quote into tasks small enough to review in one sitting, then gives each task a ceiling in turns. For headless runs the ceiling is a flag, and the JSON output reports what the run cost:

claude -p "Fix the mobile nav overlap on /services; run the gates" \
  --max-turns 12 --output-format json > run.json
jq '{turns: .num_turns, cost: .total_cost_usd}' run.json

In an interactive session, /cost shows the running total on API billing. On a subscription plan the limit is usage allowance rather than an invoice, but the same discipline applies: a task that eats a third of the week's allowance has a cost even if no bill arrives.

Match the model to the job

Not every task needs the largest model. Copy edits, alt text, redirect lists and renames across a dozen files are mechanical, and a smaller, faster model handles them well. Debugging an intermittent form failure or planning a migration is where the larger model pays for itself, because one correct answer is cheaper than five plausible ones. Switch with /model per task, and write the default for each task type into the project plan so the choice is not made fresh every morning.

Keep the context small on purpose

Most savings come from habits that cost nothing:

npm run build 2>&1 | tail -n 40
npm test 2>&1 | grep -E "FAIL|Error|expected" | head -n 30

Write a stop rule before you need one

The single most expensive pattern is an agent looping on a problem it cannot see. Jason's rule: if the same check fails three times with no new information in the output, the agent stops and a person looks. Usually the cause is outside the repository, such as an expired credential, a missing environment variable or a third-party outage, and no amount of editing will fix it. The Agentix note on where AI agents help and fail for West Palm Beach and Jupiter service businesses makes the same point about customer-facing agents: knowing when to hand off is a feature, not a failure.

Put the agent line into the quote

Suede treats agent spend as an internal line in every fixed-fee estimate, next to hosting and review time. Then log actuals per task in a plain file: task, model, turns, cost, outcome. After a handful of projects the log tells you which task types run over, and the next quote gets more accurate. A scope change reopens the estimate for the tasks it touches, not just the hours.

Small, well-defined fixes are where tight budgets work best. The Suede Scan note on five website fixes for West Palm Beach and Jupiter businesses lists the kind of change that fits in a dozen turns with room to spare.

What the budget does not cover

Token spend is usually the smaller cost. Review is the larger one, and it is human time. A team that cuts agent spend by skipping the live check has saved the cheap thing and risked the expensive one. Jason Colapietro is firm on this: the budget exists to keep review time protected, which is why note 3's review process stays in every plan regardless of how the token line looks.

Who wrote this, and why it is written down

Jason Colapietro runs the Suede AI estate with coding agents doing most of the typing, and wrote the long versions of these habits in The Claude Code Bible and Codex in Production. Suede AI, formerly Suede Labs, is a member of Claude for Startups, Anthropic's startup program. Suede AI ranks among Clutch's Top 30 AI Consultants in Florida and Top 50 AI Consultants nationwide (November 2026).

Next step

Pick your last fixed-fee project and estimate what the agent work cost per task. If you cannot, start the log this week. If you want a second opinion on the plan, the Suede AI assessment from Suede AI's West Palm Beach web practice reviews your agent setup and your live site together. For more on Jason Colapietro, see the founder page.

Built by Suede AI in West Palm Beach, Florida

Claude Code Field Notes is written by Jason Colapietro, Founder and CEO of Suede AI, an agentic platform company founded in 2024 in West Palm Beach. Suede AI's West Palm Beach SEO, GEO and web practice builds and maintains sites for businesses across Palm Beach County, from Jupiter to Boca Raton.

Suede AI ranks among Clutch's Top 30 AI Consultants in Florida and Top 50 AI Consultants nationwide (November 2026).

More from Suede AI