Got paged at 3am. Again. So I built Quell — an open-source AI agent that investigates production incidents in a Docker sandbox and ships you the root cause before you wake up. → Reads logs, git history, code → Spawns subagents for parallel investigation → Posts root cause to Slack with evidence → Runs on Claude / GPT / local Ollama (LiteLLM) Read-only sandbox. Your code is safe. MIT licensed. GitHub: github.com/Bhartiyaanshul/qu…

I'm confused, should I shift to codex or just stay with claude code?

Sunday spent well Woke up at 11 Went for pickle ball match Build something with friends Had a call with a founder for App development This is productive, not 24 hrs into building stuff. unpopular opinion.

Sunday spent well Woke up at 11 Went for pickle ball match Build something with friends Had a call with a founder for App development This is productive, not 24 hrs into building stuff.

Good Morning ☕️ Was not active for few days and can see so much new people new stuff I want build a mobile application but not able to figure out the idea, so if you guys have a idea and want to build together Let's do it

Good Morning ☕️ Was not active for few days and can see so much new people new stuff I want build a mobile application but not able to figure out the idea, so if you guys have a idea and want to build together Let's do it DM me

If you're building LLM agents, steal this pattern: Don't put your tool catalogue in the system prompt. Inject it dynamically based on which subagent is running. A "DB analyzer" subagent doesn't need HTTP probe. A "code reader" doesn't need git diff. Smaller tool set = fewer hallucinations.

2 hours yesterday on the landing page. 1 hours today. 1 hour debating between orange and amber. Building landing page is easy but design thinking is difficult i should be coding the actual product. but a tool that 'investigates incidents while you sleep' needs to look like it actually does.

I almost gave up on Quell at week 3. Docker sandbox kept failing health checks. LLM kept hallucinating tool calls. SQLAlchemy async was fighting me. The only thing that worked: shipping ugly versions to a private repo, testing on real incidents from my own side projects. Theory < production pressure.

Every "AI on-call engineer" startup is solving the wrong problem. They're building chatbots that summarize alerts. What you actually want: an agent that has already investigated by the time you read the alert. That's the bar. Anything less is just noise.

prompting claude with "is my multi-agent incident response system a good idea?" getting a 4-paragraph essay about your "massive devops TAM" and "AI-native disruption potential" vibe-coding it in 72 hours watching 12 github stars trickle in over 3 months

Most LLM agents are dumber than a senior engineer because they lack institutional knowledge. Quell solves this with skills — markdown runbooks with YAML frontmatter that auto-load when patterns match. ```yaml applicable_when: - error_contains: "ENOSPC" ``` Now your agent knows how YOUR team handles disk-full.

Most LLM agents are dumber than a senior engineer because they lack institutional knowledge. Quell solves this with skills — markdown runbooks with YAML frontmatter that auto-load when patterns match. ```yaml applicable_when: - error_contains: "ENOSPC" ``` Now your agent knows how YOUR team handles disk-full.