Skip to content

Agent and Harness Efficiency

Medium impactLive published free item

Three Codex releases in a fortnight all point at context budgets

Subagent budget accounting, three-tier context management, and the end of summary compaction. Read together, the harness is moving from hiding context cost to exposing it.

Published 11 Sept 2026, 07:45Verified 11 Sept 2026, 07:45Sources 0

What changed

Between 30 August and 5 September, Codex shipped subagent budget accounting in 0.151, replaced summary compaction with fresh context windows, and added three-tier context management with an experimental token-budget mode in 0.153. GPT-6 Astra separately added note-preserving context that Codex can search across prior windows.

Why it matters

No single release here is dramatic. Together they are a direction: the harness is being rebuilt so context spend is visible and capped rather than silently compressed. That is a bet that operators want to control context rather than have it managed for them. If it holds, the skill that matters shifts from writing good prompts to setting good budgets, and the operators who learn it first get materially more work from the same plan.

Take

Subagent budget accounting, three-tier context management, and the end of summary compaction. Read together, the harness is moving from hiding context cost to exposing it.

What to do

Turn on whatever budget visibility your harness now offers, even if you do not act on it yet. One week of watching where context actually goes is worth more than any guide, this one included.

Evidence

No public sources are linked yet.