70% of coding agent token burn is watching compiler output — JetBrains' fix analyzed

AbleBranch6 · reddit · 2026-11-11

The problem: frontier models doing chores

Breaking down an autonomous agent's step logs on a routine bug fix ($5 total): steps 1–12 (searching, tracing logic) cost $1.10, while steps 13–42 (editing, builds, test reruns, fixing imports) cost $3.80. Over 70% of the bill is paying peak frontier rates (Opus 5.5, Grok 4.7) to watch compiler output and rerun tests — work a linter plus a cheap model could do.

Junie Eco Mode: Architect + Builder

JetBrains' Eco Mode (GA in Junie CLI 26.10.20 and IDE betas) uses a single one-way handoff:

Architect (frontier model): 15 steps, strictly read-only (only write allowed: a minimal repro test), outputs a 6-point markdown spec (Goal, Scope, Where, Approach, Tests, Risks), then stops completely.
Builder (low-cost model, currently GPT-6 Luna): takes full context plus the structured plan, edits code, runs builds, parses stack traces, iterates until tests pass.

Unlike older "cheap model calls a smarter friend when stuck" multi-agent setups — where each escalation re-serializes 50k+ tokens of failed bash logs — the single handoff avoids context explosion.

Benchmarks

Broad private repos: cost per task $2.61 → $1.11 (-57.3%), resolve rate down only 0.6%.
Massive enterprise repos: $2.51 → $1.30 (-48.1%), identical 32.1% resolve rate.
Opus 5.5 max effort: $4.91 → $1.81 (-63%), resolve rate up 1 point (55%→56%) — writing an explicit plan first may prevent hallucinated rabbit holes.

Caveat

If the Architect botches its plan in those 15 steps, the Builder faithfully implements the wrong solution — there's no backward loop to re-engage the expensive model.

Original post →

70% of coding agent token burn is watching compiler output — JetBrains' fix analyzed

More from coding & agent