FROM OUR PARTNER
Elon's Cooking Up Something Big
Love him or hate him, Musk moves markets. His next launch hits July 22, and the smart money is already positioning. Our analyst found 3 stocks set to ride it — with entry points and a buy/sell playbook.
TODAY AI BRIEF
3 stories worth your next 90 seconds
1. Anthropic adds two eval commands to Claude Code.
An eval is a test that shows whether an AI feature actually works, and the new /claude-api build-eval command builds one by interviewing you and sampling production data and bug reports. The /claude-api hillclimb command then improves results by making one targeted change per round and reverting it if a held-back test set does not improve. You approve the inputs and grading at checkpoints, so the result is only as good as your samples. Read Anthropic's walkthrough.
2. Nvidia offers OpenShell, a sandbox that limits what AI agents can touch.
OpenShell is an open-source runtime that runs agents inside sandboxes with kernel-level isolation, so the limits are enforced by the system instead of the model's good behavior. It works with coding agents such as Claude Code and GitHub Copilot CLI. The documentation lists version 0.1.2, so treat it as early. Read the OpenShell documentation.
3. Fireworks releases Ember-1, a Kimi K3 variant that claims 40% fewer tokens.
Fireworks says Ember-1, built by post-training Moonshot AI's Kimi K3, gives comparable answers with about 40% fewer tokens, which matters because coding agents re-read their own reasoning every turn. Its tests show a small dip on SWE-bench Verified (92.2% versus 93.2%) and a gain on Terminal-Bench 2.1 (82.0% versus 80.9%). All figures come from the company and have not been independently checked. RuntimeWire has the numbers.
AI FORWORD
AI made the work fast. Checking it is now the job.
The speed boost. Engineer Gergely Orosz told a WorkOS event that early AI adopters once finished day-long tasks in hours. That edge faded as the tools became standard, and nearly every team now has them.
The missing payoff. By his account, whole teams now work 10 to 25 times faster, yet that has not shown up as measurable revenue growth. Instead the speed funds bigger bets, such as platform rewrites that once seemed unrealistic, and he expects platform engineers to stay essential.
The bottleneck. When building gets cheap, knowing whether the result is good becomes the scarce skill. That is the gap Anthropic's eval commands target, turning bug reports and production data into tests you can run again and again.
The fence. The other half is deciding what an agent may do. OpenShell enforces filesystem, network and process limits outside the model, which is a different promise from asking an agent nicely. The likely lesson is that the valuable work is shifting from producing output to defining what good looks like and what is allowed.
PRESENTED BY :
AI can build faster. Can your team decide better?
AI can draft the PRD and prototype the idea. Jira Product Discovery helps teams decide whether it belongs on the roadmap. Bring feedback and ideas together, prioritize as a team, and keep your roadmap connected to delivery in Jira.
SOCIAL SNAPSHOT
What people are actually watching and sharing
Search for agents. Jevgrep lets a coding agent ask plain-English questions about a codebase, such as how telemetry events are recorded, and get back files and code excerpts with line numbers. It uses the MIT license and has about 1,800 stars on GitHub.
An agent without an editor. Cline for Desktop runs the open-source Cline coding agent as a native app on macOS, Windows and Linux, with no code editor required. It is in beta, and you can bring your own API key or sign in with a Cline account.
Price check. Fireworks lists Ember-1 at $3 per million input tokens, $0.30 per million cached input tokens and $15 per million output tokens on its serverless platform.
Rules in a file. OpenShell keeps its network rules in YAML policy files that can be reloaded while it runs, and its documentation says security policies can be version-controlled for compliance work.
PROMPT STATION
Act as a skeptical senior reviewer who has seen AI make confident mistakes.
WORK TO REVIEW: [PASTE THE AI-GENERATED WORK HERE]
WHAT IT WAS SUPPOSED TO ACCOMPLISH: [THE GOAL]
WHO WILL RELY ON IT: [AUDIENCE OR DECISION-MAKER]
Do not rewrite the work. Instead give me a review checklist with these parts: the five claims or facts most worth verifying, a quick way to check each one, anything that sounds confident but has no source, what would make you reject this outright, and three questions I should ask before I rely on it.
Keep it under 300 words. If the goal is unclear, ask me before guessing.