Welcome back to Agentic Coding Weekly. Here are the updates on agentic coding tools, models, and workflows for the week of Sep 13 - 19, 2026.
1. Tooling and Model Updates
Claude Code (v2.1.277) now reads AGENTS.md if there is no CLAUDE.md file in the repo.
ZCode, official agent harness for GLM models made by Z AI, was caught silently uploading entire git history to the cloud. Z AI later made a statement that this is now fixed, all uploaded data was "destroyed immediately", and then open-sourced ZCode.
If I had a nickel for every time an agent harness was caught uploading entire git history to cloud and then later open-sourced the harness for damage-control, I'd have two nickels, but it's weird that it has happened twice.
Xiaomi published live post-training dashboard for Mimo 2.6. Pretty cool.
What I don't understand is that they are tracking benchmark scores like DeepSWE v1.1, and other benchmarks for each training step. So, if those scores are being used for hyperparameters like stopping criteria or adjusting learning rate or something, then isn't that basically benchmaxxing? They are smart people though, so I am sure they are not doing that.
Jev
Jev was the big release last week. If you are familiar with text classification using the BERT family of models or using LLMs with structured outputs, then this is something very similar. Of course, there are tons of differences in the architecture and behind the scenes of how it is designed, how it is trained, and everything.
Using Jev for text classification has the same benefits as using LLMs for text classification. The added benefit is that it's extremely fast because it doesn't need to run multiple decode steps because it's not generating text.
It's a general-purpose classifier that doesn't require fine-tuning to work reasonably well in a domain
The BERT family of models typically had a 512 input token limit, and recently, a couple of years ago, we got ModernBERT, which had an 8k tokens context window. Jev has 32k context window
With the BERT family of models, you’d typically need to decide the list of classes beforehand. To add a new class, you’d need to do a fine-tune again. Jev doesn't need a separate fine-tune for that cause you just provide your list of potential outputs or potential classes along with the input, and it just picks one of them
The limitations are that Jev still has a 32K context window, and it can only select from 10 choices or classes.
So overall, it's still lacking if you just want to do text classification, unless latency is a core constraint. LLMs have much bigger context windows, and you can pick much more output classes.
Now the thing is you can decompose most tasks into a series of decisions. The decode step in LLM repeatedly chooses the most likely token from the list of all tokens that they have in the vocabulary. So, people have been using Jev to do a lot of fun things, including generating text, controlling the computer using browser use, and a bunch of other interesting applications.
Coding Benchmarks
No major changes in terms of benchmarks last week. Here's the current state of coding benchmarks:
Model | DeepSWE 1.1 | Frontier Code 1.1 Main | Terminal-Bench 4.0 | Pricing |
|---|---|---|---|---|
DeepSeek V4.1 Flash | 74% | - | 31.2% | $0.15 / $0.6 |
GPT 6 Astra | 74% | 53.3% | 57.9% | $10 / $50 |
Fable 5.1 | - | 50.9%. | 55.8% | $10 / $50 |
Opus 5 | 74% | 53.4% | 52.3% | $5 / $25 |
GPT 5.6 Sol | 73% | 47.5% | 37.3% | $4 / $20 |
Kimi K3 | 69% | 44.2% | - | $3 / $15 |
2. Open Source Corner
stagehand - alternative to playwright for browser automation, 2x faster and 80% more token efficient
security-audit-skill - agent skill from cloudflare
3. Reading List
the senior engineer death spiral - (bonus) not about agentic coding
That’s it for this week. I’ll be back next Monday with the latest agentic coding updates.
— Prashant