Issue · 2026-08-14
ORCHESTRATION DIGEST DIGEST MODELS ORCHESTRATION DIGEST
Get tomorrow’s issue
Most AI news is noise.Here is the rest.
DeepSeek open-sourced its harness
- DeepSeek Harness v0.1 is in developer preview and open-sourced under MIT, built on the Cordis meta-framework (deepseek_ai).
- Early read: it's a React web app where everything — agent loops, models, tools, sessions — is a plugin, which is the real differentiator (Hesamation).
Gemini 3.7 Flash: the workhorse update
- Gemini 3.7 Flash is out, three weeks after 3.6, billed as Google's most intelligent model yet for coding and agents (GoogleAIStudio).
- It gained 4 points on the Intelligence Index and sits on the Intelligence-vs-Time Pareto frontier — the third Flash release in three months (ArtificialAnlys).
- Planning signal: the "introductory" price doubles on Dec 31 (simonw).
Grok 4.6 field reports
- A full day of real work on Grok 4.6 as the orchestrator agent that manages all other agents — an honest, setup-heavy review (kunchenguid).
- A heavy user's read on where 4.6 genuinely stands in the frontier race (theo).
Opus 5: the regression keeps compounding
- Cross-source and convergent: Opus 5 tops benchmarks yet feels worst in use (0xIlyy); another comparison finds it less smart, slower and more expensive than 5.6 Sol (NorthstarBrain).
- The cost angle: Opus 5 overengineers your codebase, then charges you again to clean up the slop (RylanSchaeffer).
- The counterpoint is actionable: a concrete fix for Opus 5's drift, slowness and laziness, applicable to 5.6 Sol as well (bil0090).
The harness is the product
- Uncle Bob: wrangling capricious agents into a productive harness is the software-engineering challenge of the decade (unclebobmartin).
- Anthropic engineers: half your harness fixes problems the model stopped having — a pruning principle for your own setup (0xDepressionn).
- A warning story: a flawless demo hallucinated a file path and burned $140 in a runaway loop 48 hours later — prompts were fine, the harness was not (zebratechieszts).
- Pi's own engineer on compaction: how the context window degrades output quality and how Pi implements the fix (pidotdev).
Skills and stacks to copy
- "File over agent": a free Obsidian skill with 45 commands, 1,374 forks in seven weeks (seeconvm).
- A skill that lets Claude Code watch and analyze any video — UGC ads, screen recordings, creative teardowns (mikefutia).
- A concrete local stack: GLM 5.2 and DeepSeek V4 Flash on DGX Sparks, Qwen 3.8 on a 98GB Mac Studio (JASONMCNAB).