Building with AI agents @ MythyaVerse. Agentic work, AI-first teams, everyday AI.
Lucknow
Joined February 2014
- Tweets993
- Following1K
- Followers203
- Likes2.9K
Pinned Tweet
I build with AI agents at MythyaVerse. Here I post how.
The agents aren't the slow part. Review is. So Claude Code plans, Codex does the work, and a different model checks what it wrote.
Expect how-tos on agent access and keys. And what broke.
your agent saves what it learns from your data as a Python function, so the next session reuses it and doesn't write fresh SQL.
it checks a claim before it saves it. your own sign-off is optional. i'd turn it on if a wrong number would hurt you.
github.com/Intelligible/elbi
i'd sort every agent tool into 3 bins before it runs.
reads: let them go.
writes you can undo, like a commit on a branch: let them go, then read the diff.
writes you can't undo, like an email, a deploy or DROP TABLE: it has to ask.
in Claude Code, those go in ask rules.
One desktop app for all your coding agents.
It runs Codex, Claude Code, Gemini CLI and OpenCode, keeps their sessions in one list, and shows each diff. You can pick up from your phone too.
That view has no diff review yet. It's v0.7.0, a month old.
github.com/cachix/agentaps
A local router for Codex and Claude Code.
I'd read the limits first. Reading code can go to a cheaper model, hard fixes stay on your main one, and each session shows its cost.
For Claude, the README says task matching and scoring may be inaccurate.
github.com/higress-group/HiR…
A git hook that quizzes you on your own diff before you push.
It asks 5 questions, grades you out of 100, and holds the push if you're under your team's pass mark.
Catch: the offline demo only checks answer length. Use Claude Code or Codex.
github.com/guilyx/grip
Run an agent where it can't read your SSH keys.
confine runs Claude Code in a copy-on-write fork of your repo, then fsdiff shows what changed as a patch for Git.
Catch: the network isn't limited. It can still send out what it can read.
github.com/penberg/agentutil…
Labels the phases of a long agent eval transcript and shows where your LLM judges disagree.
It's from UK AISI and Meridian Labs, and the sample report in its README flags judge agreement at 0.42. Good. You see which labels you shouldn't trust.
github.com/AI-Safety-Institu…
Checks your coding agent's tool calls against a policy file before they run, blocks .env reads and scans for 70+ secret patterns. README says no account and no telemetry.
Hooks load at session start. Open sessions aren't guarded. It's v0.0.7.
github.com/solongate/agent-s…
Checks each tool call your agent makes before it runs, then signs it into a hash chain. Edit the log later and verifying the run fails at that record.
Read its limits first. In dev mode, an agent running as you can beat it. It's a month old.
github.com/Cygnux-Labs/Trace…
Mezmo's SRE agent platform, in Rust. A team of agents reads your logs, metrics and deploys to work an incident.
Two catches: the CLI sends usage data to Mezmo unless you set DO_NOT_TRACK=1, and env vars you put in prompt fields go to the model.
github.com/mezmo/aura
Self-hosts the OpenAI Agents API. Each session runs Codex, Claude Code or MiniMax Code on microsandbox, E2B, Docker or your machine.
Read its capability table first. None of the three can run with sandbox network restricted or off. And it's at v0.0.12.
github.com/MiniMax-AI/OpenAg…
Runs a local model on your Mac behind OpenAI and Anthropic APIs, so Claude Code or Codex can use it.
It's picky. You need an M3 or newer, macOS 26.4 or later, and 36 GB for the 4-bit examples, and it only runs models from a short list.
github.com/incoai/splash
Commands in Claude Code's sandbox can still read ~/.ssh and ~/.aws by default.
Default reads cover most of the machine, including those credential files. The docs say so.
Fix? Add sandbox.credentials with mode deny. Now the OS blocks them.
the line from Claude Code's sandboxing docs: default reads cover "most of the machine, including credential files such as ~/.ssh and ~/.aws/credentials."
protect them with sandbox.credentials mode deny. link below.
code.claude.com/docs/en/sand…
Your Claude Code deny rule for .env doesn't stop every read of your keys.
Read(./.env) blocks Claude's Read tool and cat. But a Python script that opens the file itself isn't covered, and the docs say so.
Fix? Keep the rule. Run /sandbox. Now the OS blocks it.
the line from Claude Code's permissions docs: deny rules don't apply to "arbitrary subprocesses that read or write files indirectly, like a Python or Node script that opens files itself."
code.claude.com/docs/en/perm…
I build AI agents and RAG for Indian enterprises. Here I post how.
The agents aren't the slow part. Review is. So Claude Code plans, Codex does the work, and a different model checks what it wrote.
Expect how-tos on agent access and keys. And what broke.
A Zig diff viewer beside your coding agent, so you review edits with Vim keys and point at exact lines.
lgtm risk fails CI if the agent deleted a test or added a skip, but fewer asserts only count with --strict. It's one dev, 171 of 176 commits.
github.com/kunkka19xx/lgtm
A Ruby CLI for benchmarking coding agents: an instruction, a Docker image and a test script per task.
You can restore tests from before the agent ran, so it can't edit its way to a pass. Catch: on Docker, the allowlist only covers tools using http_proxy.
github.com/rails/lemans
