Kru (building with Codex) (@krutosh) header
Kru (building with Codex) (@krutosh) profile picture

I figure out Claude so you don't have to. claude code, prompts, and the workflows I actually use to ship real stuff. new tips most days 🇦🇺

Australia
Joined June 2024
Microsoft Edge now lets live sites trial AI that runs on the visitor's device. A useful test: summarise a draft without sending it to a server. If their device can't run it, ask before falling back to the cloud. blogs.windows.com/msedgedev/…
32
Long prompts can slow replies when both share the same GPUs. AWS now supports NVIDIA Dynamo on SageMaker HyperPod, so reading prompts and generating replies can run on separate GPU pools. That lets teams scale the part that's actually backed up. aws.amazon.com/blogs/machine…
1
20
Who owns the login for a job that runs at 3am? LangSmith's Tool Gateway beta lets a scheduled agent use its own credentials, while an assistant can act under the employee's. That way a nightly stock check needn't depend on someone staying employed. langchain.com/blog/langsmith…
1
1
21
Two releases of Microsoft Agent Framework on PyPI, 1.26.1 and 1.26.2, were removed after credential-stealing code was found. It could run at Python startup, even without importing the package. Agent security starts before the first prompt. github.com/advisories/GHSA-7…
21
Microsoft is rolling out Agent 365 Business Analytics in Frontier Public to connect AI spend with work completed, not just tokens. Useful direction for adoption: measure task quality and rework alongside estimated hours saved. microsoft.com/en-us/copilot/…
1
67
GitHub's Agent Host Protocol lets coding-agent sessions continue across VS Code, the Copilot app and phones. The open protocol supports shared session state across clients, with libraries in six languages. github.blog/engineering/engi…
64
Kru (building with Codex) retweeted
MAI Realtime 1 outperforms GPT-Realtime 2.1 and starts at just $0.01 per audio minute. msft.it/6018xEmwC
25
27
6
338
15,725
OpenAI added two-hour and eight-hour background completion windows to the Responses API public beta. Model token rates are 30% below Standard for the two-hour window and 40% below for eight hours. developers.openai.com/api/do…
23
Anthropic's new guide to always-on agents centers on five primitives: event inbox, timers, memory, permissions and outbox. Agents sleep between runs; durable state carries the work forward. claude.com/resources/article…
11
A new preprint caught my attention: coding agents resisted forged records when reading them directly, but made more mistakes after delegating to weaker models. This was a controlled document test, not a production failure rate. The handoff matters. arxiv.org/abs/2611.17520
14
Anthropic is cutting its disclosure window for high and critical bugs from 90 to 30 days for reports made from 1 January 2027. It cites faster AI-driven rediscovery. In-progress reports keep their existing deadlines. anthropic.com/research/coord…
10
More AI agents isn't always better. Google reports tightly coupled coding tasks worked best with sequential implementation and revision, not parallel splits. Before adding agents, check whether the work can actually be separated. research.google/blog/advanci…
2
2
86
Kru (building with Codex) retweeted
A shared AI workflow can be more useful than another custom agent. Tailscale’s Aperture now lets teams share skills with compatible agents, control access by identity and review updates. A quoting checklist could be a simple place to start. It requires an Aperture gateway, and shared instructions still need testing. The guide covers compatibility: tailscale.com/blog/skill-sha…
1
10
Building voice AI for Australian customers? Deepgram's new Nova-3 AU/NZ model is live for batch and streaming. Existing Nova-3 requests using en-AU or en-NZ use it automatically. Recheck names, addresses and noisy calls after the upgrade. developers.deepgram.com/chan…
17
Kru (building with Codex) retweeted
Your team’s AI agents could share a home for what they build. Automattic’s new Spacefast supplies hosting and a database while people keep their preferred AI. Colleagues can comment directly on published pages, then ask an agent to make changes. We like that feedback loop for prototypes. Built-in app authentication is still coming. prnewswire.com/news-releases…
1
1
15
Kru (building with Codex) retweeted
We’re optimistic about AI because we can see more people getting a chance to build something of their own. That optimism needs some care. If you’ve recently been made redundant, talk about an exciting future can feel distant. Losing an income, a team and a sense of direction hurts. You deserve time to work out what comes next, without pressure to turn it into an inspirational story. For engineers who want to try something independent, though, there is reason for hope. AI can help you explore unfamiliar tools, put together a prototype and test an idea with less work upfront. An idea you kept putting off may now be practical enough to investigate. Your experience matters here. You’ve seen where systems break, where people waste time and which problems keep coming back. Knowing what needs fixing, asking good questions and recognising a bad answer are valuable skills when software becomes easier to produce. A first step can be modest. Talk to someone about a problem you understand. Make something small that helps. See whether they use it, what they struggle with and whether solving it is worth paying for. Keep the time and money you risk within limits you can afford. Building a business still takes patience, customers and plenty of work that AI cannot do for you. Finding another role or taking time to recover are equally worthwhile paths. Our hope is that these tools give more engineers the freedom to explore an idea on their own terms. If that’s a path you want to try, your years of experience give you somewhere meaningful to begin.
1
1
18
Kru (building with Codex) retweeted
With the recent wave of vulnerability disclosures, it’s a good time to take a closer look at your own code. Turn your coding agent into a team of security auditors with the open-source skill that seeded Cloudflare’s vulnerability discovery harness. The 6 phase approach: 1. Map the codebase 2. Hunt for vulnerabilities 3. Challenge each candidate 4. Record the findings 5. Verify the findings 6. Report what’s confirmed, rejected, or still needs validation Start your hunt here: github.com/cloudflare/securi…
236
1,580
294
18,149
2,459,382