Everyone talks about prompting Claude well. Fewer people talk about scoping it well. A sharp prompt gets a good demo. It does not decide what happens when the agent is wrong, in a hurry, or slightly off about what you meant. Scope does that job: which files it can touch, which tools it can call, which actions go out without a person checking first. Claude Code permissions let you say an agent can read a repo but never touch .env, run tests but never push, open a pull request but never merge. That boundary still holds on the day the prompt did not anticipate the situation. If you only fix the wording, you get a better demo. If you fix the scope, you get a system you can leave alone overnight. #ClaudeCode #AIEngineering #AIAgents

Antigravity SDK September 28th 2026 weekly update for v0.1.20: • Multi-Backend Parity: LocalOpenAIAgentConfig and LiteRTAgentConfig now support identical policies, session continuation modes, and budget controls as local agents. • Authorization Callbacks: policy.ask_user now forwards evaluation justification strings directly into user confirmation callbacks for contextual operational approval. • Workspace Boundaries: The runtime now automatically relaxes path containment when global allow policies run without explicit workspace constraints. • Toolset Pruning: BuiltinTools.deprecated now moves directory listing and search tools into an opt-in group to reduce prompt overhead. • Resilient Tool Execution: The runtime now records failed tool executions as structured events rather than text, preserving sibling tool calls.

Some teams treat a project CLAUDE.md or Agents.md as a one time template or a default file created with init. Something goes wrong, you add a line. Claude misses a convention, you add a line. Nobody ever deletes one. Six months in, the file is 400 lines and the model is following it *less* than it did at line 80. Here's what changed how I think about it: CLAUDE.md is a scalability lever and a token optimization tool. One file. Loaded on every session. Applied by every engineer on the repo. You write a convention once and stop re-explaining your stack in chat after chat after chat. That's also why length matters so much. Everything in that file is paid for on every single session — so a bloated file isn't just untidy, it's expensive and it dilutes the instructions that actually matter. Six practices in the carousel. The one most teams get wrong is #3: not everything belongs in CLAUDE.md. Some of it is a rule, a skill, or a hook. What's in your CLAUDE.md that shouldn't be? #ClaudeCode #AIEngineering #ContextEngineering #SpecDrivenDevelopment #TechnicalProgramManagement #AgenticAI

A subagent is 97% not your task. Measured on real Claude Code transcripts: a subagent cold start is around 37,000 tokens, and only 950 of them — about 3% — are the actual task. The rest is system prompt, tool schemas, project rules, environment. Roughly 80% of that static prefix gets rewritten on every single spawn, and 53% of spawns never hit a warm cache at all. The preamble never amortizes either. It holds 25% of the bill whether the agent lives five turns or eighty, because it is re-read on every turn. Anthropic quietly moved its own number too: from "15x the tokens, but +90.2% quality" in mid-2025 to "3-10x tokens versus a single agent" plus a blunt "start with the simplest thing" in 2026. Fan-out still wins in one clear case: when a subtask generates piles of noisy output you want kept out of the main context. Split on context boundaries, not on the bullet points of your plan. Parallelism does not buy you time. It buys you a quiet context — and you pay for it in prefix. https://lnkd.in/eeVjGF3j How do you decide when a task is worth a separate agent? #AIAgents #ClaudeCode #LLM #DeveloperTools #ContextEngineering

This DX is truly a new milestone in how we review code in the AI era. Think about it... the current interface has not changed since 2011... isn't that crazy?

AI has solved how software gets written, with agents now creating upwards of 80-90% of code for advanced engineering teams. But in doing so, AI broke how software is *shipped*. With thousands of PRs coming in from a wide variety of agents, tools, and sources, how can you trust the code that has been generated? CodeRabbit has solved for this with a single platform that reviews, explains, triages, and secures every PR. At the heart of this is our Change Stack; a review workspace that turns a PR into a structured, navigable artifact you can read, question, and act on. CodeRabbit helps you understand every change so you can accurately judge code before it ships. Read more about how this works below:

Agents can create PRs in minutes, but understanding them still takes work. CodeRabbit Change Stack connects a PR's purpose, behavior, dependencies, and code so reviewers can really verify it before approving. Do you understand what you're about to merge?

PRs in Git give you timeline view, but it's so hard to understand what a PR is actually doing to your codebase. This is our way of closing the explainability gap before you merge your agentic code. Try it!

Go from reading the diff to understanding what you’re shipping. CodeRabbit Change Stack groups related changes into logical layers and ties explanations to the code.

The engineer raising the merge request isn't familiar with the code, perhaps not even the use case or initial Jira ticket these days. The reviewer most definitely isn't familiar with the code. So how do we make sense of all of this without grinding to a halt? CodeRabbit Change Stack to the rescue.