The tokens in your agent harness
Codex runs a background thread-naming call on every new session; the author reads the open-source Rust harness to show where tokens go.
AI Daily Recap
25 articles · 4 categories
The finishable daily brief
What happened in AI — Nov 1, 2026
Sunday, Nov 1, 2026
25 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Same Claude Haiku 5.5, 17 points apart: 82.0% via OpenCode vs 65.4% via Claude Code on 21 MCP servers.
- DeepSeek open-sources dsh-superagent, a router-worker coding agent, plus its architecture.
- Pi was the lightest of six coding-agent CLIs on per-job container overhead.
- SWE-bench Verified 88.6% vs 59.1% best public SWE-bench Pro score: benchmark cliff for Opus 4.8.
- Reviewing 1,700-line agent PRs needs reading order, not line-by-line diffs.
Coding-agent harness cost is the day's thread: the same model scored 82% or 65% on 21 MCP servers depending on the agent wrapped around it, and tooling to trim token spend is multiplying.
DeepSeek open-sourced a router-worker coding agent and its architecture, a direct open alternative to Claude Code.
Prefer to watch?
LLM Digest Daily Explainer: Nov 1, 2026
Harness overhead is now a measurable cost 5 items
Four posts put numbers on what coding-agent harnesses spend: tokens, container start-up time, and MCP task accuracy.
Which coding-agent CLI is best for headless workflow automation?
Six coding-agent CLIs benchmarked for per-job overhead in fresh containers: pi was lightest overall, Goose had the fastest median first response; task success is not measured.
Same model, different agent: 17 points apart on 21 MCP servers
Claude Haiku 5.5 answered 82.0% of tasks across 21 MCP servers via OpenCode but 65.4% via Claude Code; built-in tools the agent reaches for first explain the gap.
Context Mode – An Open-Source MCP Server That Reduces AI Coding Context Usage
Open-source MCP server that sandboxes tool output (claims a 98% reduction) and persists session memory across 17 platforms.
Codedict – a no-AI code index that cut my Claude Code tokens by 63%
Tree-sitter plus SQLite code index, no AI, that answers code questions up front; the author reports 63% fewer Claude Code tokens.
Agent workflows that keep humans on the decision 3 items
Teams are wrapping coding agents in review and gating loops so humans handle only the judgment calls.
How I Review 2K Lines of AI Generated Code
Reviewing a 31-file, ~1,700-line agent PR with a two-sentence description: reading order matters more than line-by-line scrutiny.
Docsferatu: An agent that drafts docs PRs while you sleep
Inngest's agent watches SDK pull requests, drafts docs in a sandbox, and opens docs PRs overnight.
Show HN: Devloops – Adversarial agent loops so humans only make the key decision
Producer, worker and gatekeeper loops let a coding agent do routine work and escalate only the key decision to a human.
Evidence over vibes: benchmarks and tool choice 3 items
Today's evals posts argue public scores and framework folklore overstate real-world results.
Is AI Cheap Enough to Replace Developers?
Claude Opus 4.8 scores 88.6% on SWE-bench Verified but the best public SWE-bench Pro score is 59.1%, and no model clears 47% on the private set.
We tested 12 search APIs for AI agents. Even the best one isn't enough
Telem.ai tested 12 search APIs for agents; a learned provider mix reached 82.6% with four providers versus 75.5% with two.
The AI cost of JavaScript frameworks
Nolan Lawson measures the AI cost of JavaScript frameworks instead of relying on claims about speed and productivity.
DeepSeek's open-source coding super-agent 4 items
DeepSeek's harness team open-sourced a router-worker coding agent and its architecture, positioned against Claude Code.
DeepSeek Open-Sources dsh-superagent, a Router-Worker Coding Agent
DeepSeek open-sources dsh-superagent, a router-worker coding agent.
DeepSeek Open Sources the Architecture Behind Its Super-Agent
DeepSeek also published the architecture behind the super-agent, not just the prototype.
QwenPU: An LLM running inside another LLM's weights
QwenPU compiles a 32-bit RISC-V CPU and vector unit into Qwen3.5-4B's weights, running a small network inside the forward pass.
An inference server for AMD GPUs, in C++ and HIP
Open C++/HIP inference server for AMD GPUs with an OpenAI-compatible HTTP API covering chat, Responses, images, audio and video.
You are caught up for this edition