Anatomy of 58.2M Tokens: Where Autonomous Coding Agents Actually Spend Money

A 14-day empirical audit of 850 model invocations across 32 workflow nodes — why unmerged retries dominate gross spend, and the four architectural changes that cut token waste by 38%.

Window: 14-Day Production Window·Empirical Sample: 850 Model Calls · 58.2M Tokens · 13 Merged PRs

1. Failure Mode Observed

While merged pull requests consumed only 586,759 net tokens on their own task rows (7.6M total), gross server spend reached 4.48M tokens per merge because post-slice developer checks resumed full CLI conversation transcripts up to 4 times per build (35M tokens) and critique loops regenerated PRDs from scratch (8.4M tokens).

2. Deterministic Architecture Fix

1) Run all deterministic static checks (completeness, test coverage, typecheck, targeted suites) before a single resumed developer session with the union of findings; 2) Cache PRD/UX/Architecture artifacts by SHA-256 brief digest; 3) Route incremental replans on diff-only critiques; 4) Enforce strict per-role token budgets.

3. Measured Production Impact

35–40% reduction in gross server spend (58.2M → ~35M/14d) and ~400k net tokens per merged PR.

Full Technical Findings & Reproducible Telemetry

1. Net Task Tokens vs. Gross Server Spend

Most industry benchmarks only report the happy-path tokens of tasks that succeeded. Over 14 days in production (850 model calls, 40.6M input tokens, 17.6M output tokens), we measure both metrics transparently: 586,759 net tokens on merged task rows versus 4,476,923 gross tokens when unmerged retries, background daemons, and planning revisions are included.

Metric                                  | 14d Baseline         | Post-Optimization Target
----------------------------------------+----------------------+---------------------------
Net tokens on merged task rows / PR     | 586,759              | ~400,000
Gross server tokens / merged PR         | 4,476,923            | <= 2,800,000 (-38%)
Developer CLI turns / merged PR         | 9.5 calls (35.0M)    | < 6.0 calls (< 2.0M/merge)
Planning tokens / merged PR (5 roles)   | 650,000 (8.4M total) | < 300,000 (SHA-256 cache)

2. Eliminating Transcript Replay in Resumed CLI Sessions

When a coding agent completes a vertical slice and spawns sequential follow-up processes for completeness, test coverage, typechecking, and linting, each resumed turn re-ingest the entire conversation transcript. Executing all deterministic static gates upfront and batching the union of failures into a single resumed turn eliminates three redundant context replays per slice.

Run these exact deterministic gates on your repository.

Start free with Brevio Core CLI or deploy the 24/7 6-Lane Harness (€0 through Mar 31, 2027).

Configure Project →All Studies