hardik verma

Staff SWE @Walmart | 10+ yrs building scalable systems | Now obsessed with Super Agents, MCPs & GenAI | I ship things that think.

[hardik verma @Hdk_verma]

Every coding agent has a sandbox now.

Almost nobody keeps it on, because it breaks npm install, git push and your dev server within the hour.

I built sandfit: it reads your Claude Code / Codex history and writes a sandbox policy that fits how you actually work. 🧵

[hardik verma @Hdk_verma]

AI models keep getting cheaper.

Enterprise AI bills keep getting bigger.

That contradiction led me to build three systems around AI agents and measure where every token actually goes.

I wrote about the experiments, failures, and architecture:

https://t.co/dtHp7gBS83

[hardik verma @Hdk_verma]

What if AI models could reason without putting their thoughts into words?

We often equate reasoning with more tokens. But a model’s internal state is numbers and vectors, not language.
Maybe the future of reasoning isn’t more tokens. It’s fewer.
Article: https://t.co/yTwij101d2

[hardik verma @Hdk_verma]

An AI agent deleted 48,000 files in 103 seconds.

It was only asked to tidy up a test folder.

@nvidia OpenShell aims to make agents safer by sandboxing them and controlling what they can access.

I wrote a beginner-friendly guide + 15-min hands-on demo.
https://t.co/euDI3MAX4S

[hardik verma @Hdk_verma]

@addyosmani @karpathy

[hardik verma @Hdk_verma]

Every coding agent has a sandbox now.

Almost nobody keeps it on, because it breaks npm install, git push and your dev server within the hour.

I built sandfit: it reads your Claude Code / Codex history and writes a sandbox policy that fits how you actually work. 🧵

[hardik verma @Hdk_verma]

Every coding agent has a sandbox now.

Almost nobody keeps it on, because it breaks npm install, git push and your dev server within the hour.

I built sandfit: it reads your Claude Code / Codex history and writes a sandbox policy that fits how you actually work. 🧵

[Chris Tate @ctatedev]

Introducing agent-desktop

agent-browser, for your whole computer

→ Computer Use for any coding agent
→ macOS, Windows and Linux
→ Runs in bg while you keep working
→ Agents use your computer in parallel, each with its own cursor

[hardik verma @Hdk_verma]

@Fahadh_4444 Yes. Docker can't run inside the sandbox (it needs the daemon), so sandfit runs it outside, still behind Claude Code's permission check. One catch: it has to be run on its own, since cd app && docker … stays sandboxed, and sandfit's forecast flags those.

[hardik verma @Hdk_verma]

@yj17469146 Yes. History never overrides the safety rules. Reads of SSH keys, cloud creds, npm tokens or .env files stay blocked even if your agent did them before, and the report flags them as ��blocked on purpose”. Same for writes to git hooks, shell rc files and the agent’s own settings.

[hardik verma @Hdk_verma]

@markfenner What it doesn’t do yet is sandbox the MCP servers themselves. They run as their own processes, so what a server reaches on the network isn’t visible in the history. It does always block writes to .mcp.json, so an agent can’t add new servers to itself.

[hardik verma @Hdk_verma]

@markfenner Both, partly. It sees every MCP tool call in the history (like mcp__github__create_issue) and lists them in the policy with how often each was used. They’re only added as Claude Code allow rules if you pass --with-permissions, since sandfit doesn’t loosen permissions by default.

[hardik verma @Hdk_verma]

Try it. It's read-only until you say apply:

npx sandfit

Works with Claude Code, Codex, Anthropic's sandbox-runtime and nono.

https://t.co/wMokGdaKLb

Tell me what it gets wrong on your history 🙏

[hardik verma @Hdk_verma]

I didn't want "trust me" numbers, so it's tested against a real sandbox (bubblewrap + Seatbelt), probe by probe:

✅ 26/26 probes behave correctly
✅ forecast matched the real sandbox on all 36 checks
✅ 96% precision/recall on commands it was never tuned on

[hardik verma @Hdk_verma]

An AI agent deleted 48,000 files in 103 seconds.

It was only asked to tidy up a test folder.

@nvidia OpenShell aims to make agents safer by sandboxing them and controlling what they can access.

I wrote a beginner-friendly guide + 15-min hands-on demo.
https://t.co/euDI3MAX4S

[hardik verma @Hdk_verma]

@addyosmani @LevelUpCoding_ @AnthropicAI @sama

[hardik verma @Hdk_verma]

What if AI models could reason without putting their thoughts into words?

We often equate reasoning with more tokens. But a model’s internal state is numbers and vectors, not language.
Maybe the future of reasoning isn’t more tokens. It’s fewer.
Article: https://t.co/yTwij101d2

[hardik verma @Hdk_verma]

AI models keep getting cheaper.

Enterprise AI bills keep getting bigger.

That contradiction led me to build three systems around AI agents and measure where every token actually goes.

I wrote about the experiments, failures, and architecture:

https://t.co/dtHp7gBS83