Categories Alternatives News Submit a Tool Advertise About
W39 Sep 14 – Sep 20, 2026
Issue W39 · Published every Monday

AI Coding Weekly

Signal, not noise. The models, tools, pricing changes and ecosystem shifts that actually matter for developers — curated once a week.

Read this week ↓ ~5 min read · 25 stories
259 pts Trust Breach ZCode silent .git upload · top HN story
800K LOC Copilot Runtime TypeScript → Rust · 128 PRs
6 models Model Sunset Retiring Oct 19 (auto-migrated)
3 tiers Auto Tiers Efficiency / Balance / Intelligence

Top Stories

What you need to know from this week
★ Editor's Pick
Editor Pick Sep 18 · Hacker News (259 pts)

A GLM Coding Agent Quietly Uploaded Developers' Full Git History

Zhipu's ZCode was found shipping encrypted full-workspace snapshots to Alibaba Cloud OSS at login — roughly 87% of the archive was .git history, and the privacy toggles people flipped did not stop it. 259 points on Hacker News, an apology and an open-source pledge the same day. The transferable lesson is not about one vendor: an agent's data boundary is now a procurement clause you have to ask about in writing, and "runs locally" is not the same as "sends nothing".

Benchmark Snapshot

Frontend Code Arena · human-preference voting

TerminalBench 2.1 — Terminal Agent Score

SWE-2 (Cognition)
92.8%
DeepSeek V4.1 Flash
90.6%
Qwen3.8 Max
86.6%
Qwen3.8-27B (local)
73.0%

Research & community-reported scores · TerminalBench 2.1

API Price — per 1M tokens (combined in + out)

Qwen3.8-Flash ¥1/¥3 per M
~$0.56
DeepSeek V4-Flash off-peak
$0.88
Gemini 3.8 Flash promo to Dec 31
$4.50
Sonnet 5 permanent $2/$10
$6.00
$0 $10

Research & community-reported prices

Models & Benchmarks

5 stories
PartForge evals Sep 18

Gemini 3.8 Flash Tops an Agentic CAD Benchmark

A third-party eval (PartForge) reports Gemini 3.8 Flash outperforming larger SOTA models at agentic CAD work — long, tool-heavy geometry tasks rather than single-file codegen. Worth reading as a hint about where the next benchmark frontier is: the tasks that separate models are increasingly the ones with a real environment attached to them.

Zhipu / vendor reporting Sep 18

GLM-5.3-FlashX Ships Into the ZCode Fallout

Zhipu announced GLM-5.3-FlashX in the same window its coding agent was under scrutiny for silent workspace uploads. Two separate matters, but a reminder to read licence and data-handling terms alongside the score: the model can be excellent and the client still be unacceptable for your codebase.

Surge Sep 16

Office-Work Post-Training Also Improves Coding

Surge trained a model on general office tasks and measured better coding performance as a side effect. It is a small result with a large implication: coding skill is not only built from code, which weakens the assumption that a specialised coding model must beat a general one on your workflow.

SSRN Sep 17

Research: Productivity Across Generations of AI Coding Tools

An SSRN paper compares productivity effects between generations of coding tools rather than between "AI and no AI" — the more useful comparison for buyers. Early read: the gains are real but concentrated in tasks with clear verification, which lines up with what we see in agent benchmarks.

AICoding tracking Sep 14-20

No New Terminal-Bench 2.1 Snapshot This Week

No comparable public score was published in Sep 14-20, so the Sep 10 table above still stands. We are flagging that explicitly instead of quietly re-printing a stale ranking as if it were news — and note that Terminal-Bench 2.1 and 4.0 are still not comparable to each other.

Tools & Editors

6 stories
GitHub Sep 16

GitHub Rewrites the Copilot Agent Runtime in Rust

A production rewrite of more than 800,000 lines across 128 PRs, moving the agent runtime off Node while keeping the JSON-RPC path working for existing clients. Large platform bets like this usually show up later as billing and governance features — which is exactly what landed the same week.

GitHub Sep 18

Copilot Code Review: Clearer Progress, Smarter Auto-Resolve

Review comments are now grouped into Open / Resolved since last review / Previously missed, comments get titles, auto-resolution understands Won't Fix and Incorrect reasons, and applying suggestions can generate a commit message. Practical effect for teams: less re-reading, fewer empty round-trips.

GitHub Sep 17

Copilot CLI 1.0.86 Reads AGENTS.md and CLAUDE.md

Custom agents can opt in (include-custom-instructions) to read repository instruction files — AGENTS.md, copilot-instructions.md and CLAUDE.md. Instruction files are quietly becoming a cross-vendor convention, which matters if you care about being able to switch tools without rewriting your setup.

Anthropic Sep 18

Claude Code 2.1.277 Adds AGENTS.md Support and Sub-Agent Isolation

When a project has no CLAUDE.md, Claude Code now falls back to AGENTS.md. Sub-agent output is delivered with an explicit "sub-agent output" header so text inside it cannot impersonate session instructions — a security fix disguised as a formatting change.

Anthropic Sep 15-18

Claude Code: Immediate Send, Remote Session Forking, Gateway Headers

A dense release run: a keybinding that interrupts the current turn and sends everything queued at once, forking of remote-control sessions into local background sessions, per-command allowed domains in sandboxed auto mode, and gateway hint headers for LLM proxies. Agent tooling is converging on the operational details enterprises asked for.

GitHub Sep 17

Ubuntu 26 GA — ubuntu-latest Migrates Oct 19 to Nov 19

GitHub Actions images move to Ubuntu 26.04 over a one-month window, so workflows that depend on system packages or tool versions should pin ubuntu-24.04 until tested. This is the kind of free-looking platform change that produces a surprise build failure in a different team's sprint.

Pricing & Business

6 stories
Hacker News / vendor statements Sep 18

The ZCode Incident Is a Procurement Clause Now

A coding agent packaging the whole workspace — .git history included — and shipping it off-device, with local toggles that did not prevent it. If you are buying agents for a team, the questions to put in writing are now: what leaves the machine, when, under what key, and can you turn it off and verify that. Ask before the security review, not after.

GitHub Sep 16

Copilot Budget Increase Requests Go GA

When a member exhausts their AI credits cap, they can now raise a request that owners and billing admins approve, adjust or reject in settings — with access restored immediately on approval. It formalises what every team was doing over chat: ad-hoc budget escalation, now with an audit trail.

Anthropic Sep 19

Claude Code Auto Mode Moves to Server-Side Classifier Billing

For API and Enterprise users plus Bedrock, Vertex, Foundry and gateway deployments, auto mode now defaults to a server-side classifier — and Anthropic states the classifier overhead is not billed, with a /status line showing whether it is running server-side. Billing mechanism, not list price, but exactly the kind of default that changes invoices quietly.

Anthropic Sep 15

Claude Code Adds Internal Markup Multipliers up to 10x

modelPricing and gateway pricing blocks now accept a multiplier above 1, capped at 10, for internal chargeback rates. For platform teams running Claude Code as an internal service, this is the missing piece for making agent usage show up as a real cost centre rather than an unmetered perk.

Velo Workspaces Sep 17

What Sandboxing an Agent in a VM Actually Costs

A write-up measuring the overhead of running coding agents inside VM sandboxes — the control that secure-coding guides keep recommending, and the one the ZCode incident makes non-optional. This is the cost line most TCO models omit entirely: isolation, snapshots and rebuild time are real money.

GitHub Sep 18

Six Copilot Models Retire Oct 19 — Budget the Migration

The deprecation list (Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Grok 4.5) matters for two reasons: anyone pinning a model in CI or a custom agent has to move, and the replacements price differently. Automatic migration is convenient, but nobody re-runs their cost model when it happens for them.

Ecosystem & Community

5 stories
The Decoder Sep 17

EU President Warns About Agents Escaping Their Environment

Commentary from the EU president that agents breaking out of their sandboxes preview what is coming. Regulatory attention arriving at agent containment rather than model capability is a signal: expect procurement forms to start asking about sandbox boundaries long before any rule takes effect.

arXiv 2609.17817 Sep 18

arXiv: Poisoning Self-Modifying AI Coding Agents

A Trusting Trust-style paper on poisoning agents that modify their own code or configuration. It is the theoretical version of the ZCode story: once an agent can change what it runs, the attack surface includes the agent itself, not just your dependencies.

Augmented SWE Sep 18

Guide: How to Secure Coding Agents

A practical write-up on securing agents — permissions, sandboxing, review gates — arriving in the same week as a headline breach and an arXiv paper. Useful as a checklist to hold vendors against, especially the parts about what an agent can read versus what it can send.

Hacker News Launch Sep 17

Skillsync (YC W26): Sessions Portable Across Agents

A Launch HN for making AI chat sessions portable between agents — 63 points and 54 comments, which is unusually engaged for this category. It is aimed squarely at the switching pain we track: teams avoid moving tools because their accumulated context cannot come along.

Hacker News Sep 17

A Curated Marketplace for Agent Skills

skillbay.sh launched as a human-curated directory of agent skills — 25 points and 18 comments. Small, but another data point that the ecosystem is formalising around reusable skills rather than per-tool prompt libraries.

Our Take

The editorial view
01

The week's story is a data boundary, not a benchmark. ZCode packaging workspace snapshots — .git history included — and shipping them off-device, while the toggles people flipped failed to stop it, turns "what does this agent send" from an architecture question into a purchasing question. We have been writing about lock-in and vendor risk on our alternatives pages for months; this is the week that made it concrete, and it is the kind of thing an official product page will never mention.

02

GitHub spent the week acting like an infrastructure company. Rewriting the Copilot runtime in Rust, shipping budget-increase requests, moving auto model selection to cost tiers, and announcing six model retirements with automatic migration all point the same way: agents are becoming metered, governed infrastructure. For buyers the actionable part is unglamorous — the Oct 19 model deadline, the Nov 2 Actions default change, the Ubuntu window, and whether your credits cap has an approval path.

03

Instruction files quietly became a standard. Claude Code now falls back to AGENTS.md, and Copilot CLI custom agents can opt in to read AGENTS.md or CLAUDE.md. That is good news for anyone worried about switching costs — and it is exactly why we have kept our own project knowledge in plain files rather than in one vendor's proprietary format. Tool portability is being standardised from below by whatever the agents agree to read.

Older issues are loaded on demand to keep this page fast.