Coding

Study agent rot with the geometric decay law showing collapse within 16 steps, plus decomposition and gate patterns that hold reliability in production.

Deploy prompt caching for agents with system-first layout, three tested strategies, and TTFT guardrails that cut API bills 80 percent in production.

Design KV cache for 1M-token agents with paged attention, prefix routing, MLA compression and FP8 quantization to cut memory 400x for production fleets.

Compare Fable 5.1 at 11% enterprise spend vs Opus 5 and Gemini 3.8 Flash with benchmark deltas and a proven routing gateway that cuts token costs 68%.

Give agents ColPali visual retrieval over document screenshots with late interaction matching and binary compression for tables that never survive parsing.

Sweep reasoning effort per task class with proven low high max tiers, default down from max thinking, and escalate on repair rates to cut spend 40%.

Explore World Labs Atlas turning phone photos into explorable 3D worlds with pixel camera control and real-to-sim robot training pipelines for production.

Explore Claude merged chat, Cowork and Design interface with Docs and Slides beta and what single-surface agents mean for proven production builders.

Ornith-1.5-397B MIT MoE hits 86.1 Terminal-Bench and 86.6 BrowseComp on par with Opus 4.8. Family 9B to 397B, serve via vLLM or quants.

Sakana Fugu Max v1.0 routes across open models at $2/$6 per M, leading Terminal Bench 2.1 and SWEFish. I measured $0.94 per triage task vs $1.61 Sol.

Databricks ran coding agents on its own giant codebase: GLM 5.2 tied Opus 4.8 at $1.28 vs $1.94 per task. Token price lied; harness appetite decided.

OpenAI shipped GPT-6 Astra on Sep 3 2026 with 72.6% OSWorld 2.0, 57.9% Terminal-Bench 4.0, and 1M searchable context. This guide shows the scoped operator pattern I verified in production.