How TaskWire cut AI agent costs by 31%

Background

The team builds several products, including TaskWire, a client web app and a restaurant point of sale system. Its agents run in Claude Code on Claude Opus 5.5.

Over seven weeks the team recorded 83 agent sessions. In 64 of them, agents worked from a TaskWire board: they claimed tasks, logged progress, handed work to each other and submitted finished tasks for review. In the other 19, agents worked in a single ongoing chat with a person. Claude Code records the tokens used by every model response, and this study priced all 83 sessions at Claude API list rates.

Cost per model call

With TaskWire, the average model call cost 10.3¢. Without it, the average was 15.0¢. Both projects that ran both ways show the same result.

Most of the cost of an agent is context: the conversation it re-reads on every call, which accounted for 56% of spend in this study. TaskWire agents averaged 293K tokens of context per call, compared with 416K without TaskWire. Because each task is a separate unit of work, agents start new work from a short context instead of the full history of the day.

Oversized sessions

The most expensive sessions are the ones that grow close to the 1M token context limit. Nine sessions passed 800K tokens. Together they cost $746, 39% of total spend, while making up 11% of sessions.

Without TaskWire, 26% of sessions reached that size. With TaskWire, 6% did.

Estimated savings

The 12,270 model calls made with TaskWire cost $1,262. At the cost per call measured without TaskWire, the same calls would have cost $1,840, a difference of about $580.

Counting only the additional context those calls would have carried, priced at the cache read rate of $0.20 per million tokens, the difference is about $300. At these rates, each additional 1,000 model calls saves roughly $47.

Human involvement

With TaskWire, agents completed 588 tasks from 388 human prompts and created 324 tasks for each other. Each prompt led to 2.0 file edits on average, compared with 1.5 without TaskWire. Responses were also shorter, with a median of 673 output tokens per call against 895, because progress updates were posted to the board.

Coordination overhead

Data returned by TaskWire to the agents, including task lists, task details, messages and attached screenshots, totaled about 7.8M tokens. That is 0.2% of all tokens in the TaskWire sessions.

Data

Methodology: Claude Code session logs for all 83 sessions with token data between September 18 and November 3, 2026. 82 sessions used Claude Opus 5.5 and one used Claude Opus 4.8. A session counts as using TaskWire if it called any TaskWire tool. Costs use Claude API list prices as of October 2026: $4 per million input tokens, $20 per million output tokens, $0.20 per million cache reads and $8 per million one hour cache writes. The team used a subscription plan during the study. The two groups worked on different tasks, so savings figures are estimates.