Nick Frolov @v1becoder

A coding agent fails for three boring reasons: - a task that is too big and poorly specified; - not enough relevant knowledge; - too much junk in context. Working with agent is a balancing act to pick and keep only the right context

OK, so all I need is M5, 64Gb of RAM, and a plane 😁

Kirill Skrygan @kskrygan

5 days ago

JetBrains takes its first step toward Local AI. MacBook M5 users can set up Junie Local in one click and get a solid agentic experience. Unlimited & Fully local. Folks also had some fun presenting it on a plane :) Check it out & stay tuned for more: junie.jetbrains.com/local

20 45 408 64K 115

JetBrains takes its first step toward Local AI. MacBook M5 users can set up Junie Local in one click and get a solid agentic experience. Unlimited & Fully local. Folks also had some fun presenting it on a plane :) Check it out & stay tuned for more: junie.jetbrains.com/local

JetBrains @jetbrains

5 days ago

Junie can now run entirely on your machine, and it’s completely free to use. One command: /local. We picked and optimized the whole stack around the model: agent harness, model parameters, and inference engine. No tokens, no quota, no code ever leaving your machine.

23 59 374 103K 112

Read up on details blog.jetbrains.com/junie/2026/08/…

Up to 2x Generation Boost: Combined Multi-Token Prediction (MTP) with n-gram speculative decoding, accepting up to 11 predicted tokens per step.

We just shipped Junie Local, running fully on-device on a MacBook M5 with Qwen 27B. Junie downloads inference engine (mlx-vlm with custom additions) and model weights, so you can use everything with just one command.

Junie now supports Github Copilot subscription, use your existing Github subscription to access models available from Github Copilot. No any extra subscription needed Activate via /account -> External API keys -> Enable Copilot

We are defaulting Gemini3.7 Flash in Junie and giving a solid discount for it. blog.jetbrains.com/junie/2026/08/…

Short summary of qwen 3.8 27b testing An old Qwen is better than two new… Well mainly faster, qwen 3.8 is significantly slower than 3.6

The new Gemini 3.6 Flash is live in AI Assistant and Junie. See how we built an objective evaluation of an engineering task to compare it against the last Flash. Try it yourself in your JetBrains IDE. @GoogleDeepMind

The observation: a cost discrepancy Our preliminary data shows a significant gap in monthly spend among heavy, all-day AI coding users: Junie users: ~$2,000/month Claude Code / Codex users: ~$5,000/month We are now running a formal study to verify these numbers. The core question is simple: Is Junie cheaper because it is genuinely more cost-efficient — or are users spending more in other tools because they feel more productive there and use them more? Our working theory is that a big part of the gap comes from inference economics, especially context management and token caching. 1. Context compaction strategy Junie aggressively auto-compacts conversation history. In practice, a Junie conversation does not keep growing indefinitely and never exceeds ~150k tokens. Other tools tend to compact only when the conversation approaches a high threshold of the available context window. That sounds fine in theory. But in real developer workflows, it can become expensive. 2. The cache problem At first glance, constant compaction looks counterintuitive: changing the context can invalidate token cache benefits. But token caches are short-lived. In many real workflows, developers switch tabs, jump between conversations, write code, review diffs, get interrupted, and come back later. By the time they send the next message, the cache may already be gone. So if the conversation history has grown very large, the user can end up repeatedly paying for a huge amount of context that is no longer cached and often no longer useful. 3. The practical result Junie’s aggressive auto-compaction may save a lot of tokens specifically because real development work is interrupted, non-linear, and multi-threaded. This is the part that matters: cost efficiency is not only about model price. It is also about how intelligently the product manages context over time. There is also a strategic angle here. Companies that monetize primarily through token consumption are not naturally incentivized to minimize token usage as aggressively as an IDE vendor trying to deliver the best AI experience at the lowest sustainable cost. That may become one of the biggest differences between standalone AI coding tools and deeply integrated AI inside IDEs.

ACP was built for exactly this: making agents portable across tools. Welcome aboard, Harbor.

Harbor Framework @harborframework

2 months ago

Harbor now natively supports evaluating ACP agents: 𝚑𝚊𝚛𝚋𝚘𝚛 𝚛𝚞𝚗 --𝚊𝚐𝚎𝚗𝚝 𝚊𝚌𝚙:𝚓𝚞𝚗𝚒𝚎 This brings 36 agents directly to Harbor, plus all future ACP agents the day they are added.

1 3 12 12K 4

AI Spec-Driven Development is like teenage sex. Everyone talks about it. Nobody really knows how to do it properly. Everyone thinks everyone else is doing it. So everyone claims they are doing it. You guys are doing it right?

Opus 4.7-optimized prompt with GPT-5.5 model will interpret the hyper-specific, step-by-step constraints as restrictive micromanagement. It will narrow its search space, slow down, and likely produce an overly mechanical answer because you didn't leave it room to find its own efficient path. If you use a GPT-5.5-optimized prompt with Opus 4.7 model will look at the broad, outcome-oriented instructions and take them completely literally. Because you didn't explicitly define the boundaries, negative constraints, or step-by-step scope, Opus 4.7 will likely give you the bare minimum of exactly what you asked for, entirely missing the unspoken nuances you actually wanted. Fun time to be in Multi-LLM agentic harness business. If this deviation will continue we will need to have model specific prompt adaptors and/or small LLM which rewrites prompts to the target LLMs

I was reading Opus 4.7 and GPT 5.5 prompting guides and very curious about the different path which they took. Opus 4.7 wants no-magic, tell it exactly what to do, what not to do, and the exact scope of the task. Which is the reason it reduced calling extra tools if you didn't ask it to. GPT-5.5. is the opposite, tell it what a successful final answer looks like and let it find the path. How would agentic harness behave if it's prompts are optimised for one or another?

Vibecoding in brave mode be like

Had a pleasure to do a podcast with Leanne Bevan from Grey Matter Discussed Agentic AI Development, Jetbrains and of course Junie Apple Podcast podcasts.apple.com/nl/podcast/gre… Spotify open.spotify.com/episode/4HjViG…

AI agents are not a substitute for bad engineering practices. AI agenting development is an amplifier. More generated code shifts the bottleneck to code reviews and testing. If your team somehow survived without proper engineering, it will not be able to keep doing the same in our AI-assisted development times. Not only context engineering is important. Engineering practices need to be back in fashion - unit tests, code coverage requirements, auto-tests, auto-builds, etc. Here is my engineering maturity list tiers