Anthropic upgrades Claude Code Projects to run parallel AI workers
Claude now scopes a request, splits it into parallel threads, reviews the outputs and assembles the result, and keeps working after you step away.
AI Daily Recap
39 articles · 5 categories
The finishable daily brief
What happened in AI — Nov 14, 2026
Saturday, Nov 14, 2026
39 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- Claude Code Projects now runs a coordinator that splits work into parallel threads and reviews the results.
- An all-agent org on Amp merged 1,119 PRs in a week and grew its own bureaucracy.
- Google: stale specs and resolved bugs in a repo read as instructions to coding agents.
- Three agents that each respect their permissions can still leak data together.
- DeepSeek opened its app's search, file, OCR and voice tools as APIs and ported DeepEP to Huawei Ascend.
Coordinating many coding agents is turning into a product feature: Claude Code Projects now scopes, delegates and reviews parallel work itself, and one founder's all-agent org merged 1,119 PRs in a week.
The day's best engineering writing put agent failures outside the model, in stale repo context, drifting job state, and per-tool permissions that leak data once agents are chained.
Prefer to watch?
LLM Digest Daily Explainer: Nov 14, 2026
Agent fleets: coordination moves into the product 4 items
The manager role in multi-agent coding is being automated: a coordinator scopes, delegates and reviews, and the human steers from one thread.
My agent org speedran a company
An all-agent engineering org on Amp merged 1,119 PRs in one week, and the agents rebuilt big-company bureaucracy when told to organize like one.
Show HN: CodeRelay – Let your coding agent delegate work to a teammate's agen
Hands work between teammates' coding agents with API contracts pinned to Git commits, so two agents stop editing the same files.
Goal Engineering
LangChain argues open-ended agents need explicit metrics to optimize, and that a badly scoped goal mostly buys wasted tokens.
Context and state: where agents actually fail 4 items
Two of the day's best writeups locate agent failures outside the model: stale repo context and drifting job state.
Agent Distraction
Google warns that obsolete specs, resolved bugs and stale docs in a repo read as instructions to a coding agent and send it on side-quests.
Why a 98%-accurate model leaves an agent's state right only 21% of the time
Per-step errors compound across a long job; the AXI paper holds job state on the host and gates every reply deterministically before applying it.
Write the eval before the fix: building a WhatsApp transcriber with LLM-TDD
A worked example of LLM-TDD: each bug in a voice-transcription pipeline gets a failing eval case before any prompt change.
Show HN: Portal2Bench – Can AI agents solve Portal 2?
An agent benchmark on Portal 2 test chambers capped at $30 per run: GPT-6.1 Sol solves 20 of 26, Opus 5.5 solves 11.
Agent permissions don't compose 4 items
Per-tool permissions are not a security boundary; the fixes on offer are reference monitors, staged effects and hard spend caps.
Transitive Policies for Agents: Agent Permissions Don't Automatically Compose
Three agents that each obey their own tool permissions can still leak customer data once an orchestrator chains them together.
The Missing Ring-0 Protections for AI Agents
Argues coding agents need a reference monitor they cannot edit, and that it only works if file and command effects are staged first.
Before I Trust an Agent, I Try to Break Its Box
A hands-on attempt to escape a per-agent sandbox from the inside, listing what broke and what is still untested.
Show HN: Shotoku – Local-first action governance and budget caps for AI agents
Local, deterministic per-vendor and per-day spend caps for agents, enforced before a payment goes out, with a tamper-evident audit log.
Harness building blocks and reference stacks 5 items
Builders shared the practical parts of a harness: tool loops, browsers, approval flows, and a catalog of what production teams run.
Build Your Own AI Coding Agent Harness
Vercel's course builds a coding-agent harness from the tool loop up, covering context pruning, sandboxing and subagent delegation.
Show HN: Browserr – An MCP-first browser for AI agents (one Rust binary)
An MCP-first browser engine in one Rust binary with no CDP or Chromium; it blocks when unsure and hands the page to a human.
Building a human-in-the-loop Slackbot agent
A reference Slackbot where code and Claude clear easy cases and people approve risky ones, built around outages, token overruns and bad output formats.
WhatWeRun, a database of real production AI agent stacks
A catalog of production agent stacks drawn only from public writeups, with costs kept on the source's own basis.
Show HN: SunsetLint – find retiring AI model IDs in your GitHub repos
Finds retiring model IDs in your repos by file and line, next to the shutdown date and the provider's suggested replacement.
Models, APIs and non-NVIDIA infra 3 items
DeepSeek shipped on two fronts: its app's tools became APIs, and its expert-parallel library now targets Huawei Ascend.
DeepSeek Opens Its App's Search, File, OCR and Voice Tools as APIs So Developers Can Rebuild the App
DeepSeek exposes its app's search, file, OCR and voice tools as APIs, so developers can rebuild the app's features on their own stack.
Training and Inference on Huawei Ascend NPUs
DeepEP-Ascend ports DeepSeek's MoE all-to-all communication library to Ascend NPUs with FP8 dispatch and a buffer API matching the NVIDIA version.
I Tried Every LLM Architecture Change Since 2020
An ablation of architecture changes since GPT-3 finds fine-grained MoE the biggest win, with linear attention and hyper-connections also paying off.
You are caught up for this edition