This isn't a niche edge case. It's the default state for uncoordinated agent systems. And unlike infrastructure waste — idle servers, over-provisioned databases — AI compute waste is invisible. There's no "wasted tokens" column in the billing dashboard. Every API call looks like intentional work, even when it's undoing, redoing, or duplicating intentional work.
Three Ways Uncoordinated Agents Burn Budget
The waste falls into three distinct categories. All three are architectural, not algorithmic — meaning a smarter prompt won't fix them. They require coordination primitives.
Duplicate processing
Two agents claim the same task from an unguarded queue. Both complete it. One result is discarded. The tokens are not.
Conflict remediation
Agent A writes a record. Agent B overwrites it. Agent A detects the inconsistency and reprocesses. Net result: the same task runs three times.
Idle context burn
Agents waiting on a blocked downstream agent retain their full context window. They poll, retry, and eventually timeout — all billable activity.
A Concrete Example
Consider a content pipeline: 10 agents each responsible for processing a batch of articles — summarizing, tagging, and publishing. No coordination layer. Here's what actually happens in a typical run:
| Activity | Intended | Actual | Waste |
|---|---|---|---|
| Articles processed | 500 | 500 | — |
| Total agent invocations | 500 | 680 | +180 (36%) |
| Conflict overwrites | 0 | 42 | 42 reruns |
| Effective cost per article | $0.004 | $0.0054 | +35% |
| Monthly cost at scale (50k articles) | $200 | $270 | $70/mo wasted |
That's a content pipeline running at modest scale. Now apply the same multiplication to a data ingestion fleet processing millions of records, or a customer support system handling thousands of tickets per day. The absolute dollar waste scales, but the percentage — 30–60% — stays consistent across architectures. We've measured it across a dozen different workloads.
Rule of thumb: If your agents share a work queue and you haven't implemented lease semantics, assume you're paying for at least 1.3x the work you actually need. For fleets with write-heavy tasks, budget for 1.5–1.6x.
How to Calculate Your Own Waste
The formula is straightforward, and you don't need special observability tooling to run it. Pull your agent execution logs for the last 30 days and calculate three numbers:
Duplication rate: Count tasks where multiple agents recorded a completion event for the same task ID within a 5-minute window. Divide by total completions. Anything above 5% is a coordination problem.
Reprocessing rate: Count tasks that were completed, then re-opened and completed again. This catches conflict-driven reruns. Above 3% is significant at scale.
Idle token rate: If you have per-execution token counts, compare tokens consumed by tasks that ultimately failed or were superseded versus tasks that produced a retained output. A ratio above 0.15 (15% of tokens in failed/discarded runs) means your retry logic needs attention.
Multiply your current monthly spend by the sum of these three rates and you have your coordination waste. Most teams are surprised by the number. Our Runaway Cost Calculator automates this if you'd rather plug in your numbers directly.
What Coordination Actually Costs to Fix
The upfront cost of adding proper coordination — lease semantics, conflict detection, budget caps — is real. If you're building it from scratch, expect two to three weeks of engineering time for a solid implementation that handles edge cases (network partitions, agent crashes mid-lease, etc.).
The payback period at a $500/month API spend and 35% waste rate is about six weeks. At $2,000/month it's under two weeks. The math gets obvious fast once you've put a number on what you're currently burning.
The alternative to building it yourself is using an orchestration layer that ships it as infrastructure. The pattern is the same either way — the question is whether it makes sense to own and maintain it, or to treat it as a platform dependency. Either path beats continuing to pay the coordination tax indefinitely.