Nathan Burg

@FlyaKiet @komran21 I’m going to try this! Testing has become my bottleneck at the moment

@komran21 directly through chrome CDP chromedevtools.github.io/devtools-proto…

Relaxing load test to fall asleep to (sound on 🔊). Automated by Codex and ran in real time. This is not sped up. I clearly have too much time on my hand waiting for these tests to run.

Most developers don't jump into agentic development. They climb there. I keep seeing the same onramp: 1. No tools. Writing everything by hand. 2. ChatGPT or Claude for single-turn Q&A. Copy-paste the answer. 3. First time in Cursor, Claude Code, Codex, etc. Single thread, single task. Watching every line. Babysitting. 4. Still single-threaded, but learning. You start to internalize the failure modes. Where it hallucinates, where it drifts, what instructions it needs. You're building the muscle of how to steer. 5. You trust yourself to let the agent finish a task on its own. Not because the agent got smarter, but because you learned how to set it up to succeed and catch it when it doesn't. 6. You realize you could be running three of these. Five. Ten. Full agentic development. The progression isn't that the agent magically gets better. It's that you develop the muscle of how to use them effectively. What they can do, what they can't, and how to set the guardrails so you can let them run. Does this match your experience? What step are you on?

[ 🄿🄰🅂🅃🄴🄳 1️⃣ 🄻🄸🄽🄴🅂 ] hate getting the text block instead of the actual text you pasted in @opencode? here's how to fix it:

Just started using Hermes agent today. I configured Hermes with Daytona sandboxes + my Codex subscription and got it running experiments in no time. Seamless setup and so fast to get working. More to share soon. Great job @NousResearch

new @GitHits_com update looks fire 🔥:

Pi support is coming soon

Are you using a coding agent that's not on this list? `npx githits@latest init` auto-detects and configures GitHits for these agents: - Claude Code - Cursor - Windsurf - VS Code / Copilot - Cline - Claude Desktop - Codex CLI - Gemini CLI - Google Antigravity - OpenCode If your tool isn't here, we want to know. Drop a comment or open an issue on our CLI repo: github.com/githits-com/gi… We're adding new agents regularly and community requests go straight to the top of the queue.

If you're on a cloud provider, use their native secret manager. AWS Secrets Manager, Azure Key Vault, Google Cloud Secret Manager. They all have one. HashiCorp Vault is solid if you want something cloud-agnostic. For one-offs where you just need to hand a .env to a teammate, something like One Time Secret works well.

how do teams usually share .env variables securely ?

Depends on the product. Go with whatever interface makes more sense for your users. If both would genuinely add value, build both. Start with a well-architected backend and implementing more than one interface becomes a lot easier. Factor in how good coding agents are getting and building out and maintaining two interfaces is genuinely doable. If the value is there, why choose?

Should solo founders build web apps or mobile apps in 2026?

@suni_code githits.com

Drop your project URL 👇🏻 Let’s drive some traffic...

We removed our waitlist! Anyone can sign up for GitHits now, and everything is free while we're in beta. npx githits@latest init That one command creates your free account, detects your coding agents, and configures them. Here's what they get: - Unified search across code, docs, and symbols in any package across 11 registries (npm, PyPI, Crates, Hex, Maven, NuGet, Go, RubyGems, Packagist, vcpkg, Zig) or any GitHub repo - Code navigation: list files, read source at exact line ranges, and deterministic grep that chains directly into reads. No cloning. - Package docs browsing across all supported registries, both hosted and repo-backed - Dependency intelligence: package overviews, dep graph analysis, vulnerability checks, and changelogs before your agent picks a dependency - Source-attributed, license-tracked code examples from open source Paid plans are coming at launch. Get in now.

Context is King no. 4 in San Francisco yesterday. A few of the highlights below, full talks on YouTube: youtube.com/playlist?list=… Next up: London. @bergr7 (@flowaicom), Itai Smith (@trychroma), @Romainsestier (@StackOneHQ), @OtsoVeistera (@thetokenco), @NathanBurg (@GitHits_com), @aiven_io

Wrong photo in the post😂. This is the actual example.

One of the things I love about being in San Francisco is getting to meet builders who are passionate and excited about the problem they’re solving. Here’s a quick example of why that matters. I wanted to understand how Superset’s desktop app handles config resolution when there are config files at multiple levels, the main repo, the worktree, and a user override, and what config.local.json is for. Using Anthropic’s Opus 4.7 without GitHits, it took 10.1k (1.8k + 8.3k) output tokens and about 7 minutes. With GitHits, it took 2.1k output tokens and 40 seconds. How do I know this? ChooChoo. @alexrigler from ChooChoo and I took a walk along the Embarcadero, and I was telling him how I’ve been using ChooChoo to evaluate token usage across coding agents, comparing sessions with and without GitHits. I mentioned it’d be really useful to compare sessions side by side, since I’m constantly running parallel benchmarks. Right after we got back from the walk, Alex messaged me: “I’m gonna try to get this working before your demo tomorrow.” By the next morning, ChooChoo had a brand new Arena feature, session comparison, shipped and working, and I was using it live onstage that night while demoing GitHits. A conversation on a walk to a production feature overnight. This is why we’re excited to be working alongside teams like ChooChoo. They’re building the observability and evals layer for coding agents, and it’s exactly the kind of tool we rely on to understand how agents perform with GitHits in the loop. When two teams are building in the same space and actually using each other’s products, things move fast.

Really excited I got to demo @GitHits_com last night at Context is King! I walked through our two core primitives: • Code Examples: GitHits searches open source code, evaluates quality signals like adoption and contributor activity, and distills it all into a single canonical example with every source cited • Code Navigation: when you know the library you’re working with, your agent can grep, glob, and read any version of any open-source repo on GitHub, plus packages from any of the 10 registries we support through our Code Index We also just removed our waitlist. Anyone can sign up now. Run `npx githits@latest init` in your terminal and you’re good to go. Check out the full demo below. Thanks to @flowaicom and @aiven_io for hosting a great event! youtu.be/aXCPSueyWvk?si…

Demoing at Context is King #4 in SF on May 5th, the meetup hosted by @flowaicom and @aiven_io. I'll be running a live demo of @GitHits_com inside my own agentic dev flow, walking through two parts: pulling canonical code examples with citations back to the source libraries, then code navigation that greps and reads through real open source projects to answer questions from the actual codebase. No slides. Happy to take live audience requests for the demo. Come check it out if you're around. Sign-up link is in the comments.

@heyblake githits.com

Fork it Drop your landing page URL I'll give 1 piece of advice to as many of you as I can

Tonight I'm hosting the 6th CodexSF meetup at @WorkOS (also streaming on X and YouTube) Stop by for some great talks, pizza and drinks! Link in first reply. Demos by: @vineet_reddy1: an autonomous genetic research system where Codex helps scientists generate research papers @rjp0x: a no-code orchestration flow: an orchestrator that plans and tests, dispatching subagents that work like a full-stack team @Nikiag65: using Codex's newly launched hooks for hands-off control @NathanBurg: parallel Codex agents across isolated git worktrees, with an adversarial plan reviewer and a shared skills directory, for drift-free autonomous runs @curious_queue: Codex hooks, a zellij setup, and a cron job in Hermes that contextually nudges Codex to keep executing

testing something new in Codex currently invite-only but @GitHits_com is *really* good it specifically deep-searches GitHub open source repositories for patterns, so your AI can learn from REAL WORLD EXAMPLES and not just docs