A Practical Guide to Reducing Token Spend

David Cramer of Sentry recently wrote about burning through $10,000 a week on AI tokens, pinning much of the blame on a code-review skill called Garfield that leans on a coordinator agent to dispatch dozens of sub-agents. Each one reviews code, reports back, gets adjudicated, then the cycle repeats. It's clever, it's genuinely hard to build, and it's also brutally expensive — every step routes through a non-deterministic model judging the output of other non-deterministic models.

So someone rebuilt Garfield as what they call a "swamp" workflow, keeping the exact same functionality but swapping the coordinator's judgment calls for regular deterministic code. The original skill chewed through 4.5 million tokens, took about 12 minutes, and spun up 23 sub-agents to review a representative problem. The swamp version did the identical job in roughly 500,000 tokens, 6.5 minutes, and just 3 agents. That's an 8x drop in token spend and half the wall-clock time, without giving up any of the review logic.

The trick isn't magic, it's discipline. Swamp stores each sub-agent's output as versioned, typed data, so later steps can see prior work instead of redoing it, and it only calls the LLM where judgment is actually required — leaving the boring, reliable parts of the pipeline (formatting checks, test verification, bookkeeping) to plain code. The write-up frames the whole exercise as a four-step process: understand what the existing skill does, translate it into a swamp extension, black-box test both versions side by side, then refine based on what breaks.

That testing step surfaced something worth noting on its own. When a bug was buried in a complex code diff, the original Garfield skill quietly declared victory and left the defect in place — it failed open. The swamp version, hitting the same budget limit, instead reported that unresolved findings remained. It failed closed. Same underlying constraint, very different behavior, and one of those is clearly the one you want in production.

The bigger argument here is that most agentic workflows today ask the LLM to do work computers have handled reliably for decades — sequencing, aggregation, pass/fail logic — and that's where the token bill actually balloons. Building a lightweight deterministic scaffold around the model, and only calling on it for the parts that genuinely need judgment, is presented as the highest-leverage change available to anyone worried about inference costs right now.