Most of my time goes into PANOPTIS (formerly AgentWatch) at Shinraya Tech Labs. It sits between an agent and its tools and decides on every call before it runs: the rm -rf gets blocked, the git push --force waits for a human, and instructions hidden in a web page get stripped before the agent ever reads them.
If you're running an agent that can touch a database or a shell, try this:
pip install panoptis # Python 3.12from panoptis.protect import ActionBlocked, protect_tool
@protect_tool
def execute_system_query(query: str) -> str:
return db.execute(query) # never reached for a blocked call
try:
execute_system_query("DROP TABLE users;")
except ActionBlocked as blocked:
print(blocked) # execute_system_query: block: [threat:SQL_DROP] DROP TABLEWant to see it stop a real attack first? Run panopti playground for a local browser playground.
The things I check on every agent I build or review. Steal it.
Open the 10-point checklist
- Scope every tool. Each tool gets an allowlist: which folders, tables, domains or APIs it can touch. "Access to the database" is not a scope.
- Put a human in front of anything you can't undo. Deletes, force-pushes, payments, emails to customers.
- Treat tool output as untrusted. A web page, a GitHub issue or a PDF can contain instructions aimed at your model. Strip them before the model reads it.
- Keep secrets out of the context window. Tools hold scoped credentials; the model never sees a raw API key.
- Set hard limits. Max steps, max cost per run, and a timeout on every tool call.
- Make retries safe. A retried call must not charge a card or send an email twice.
- Log the whole run. Input, reasoning, every tool call and its result, so you can replay a failure instead of guessing.
- Keep a test set. 20 to 50 real tasks you rerun before every prompt or model change.
- Have a kill switch. One command that stops every running agent.
- Plan for the model failing. Rate limits, outages and garbage output need a fallback, not a stack trace.
| 🧠 adacompute | Lets a model decide how much to think per input, instead of spending the same compute on an easy question and a hard one. |
| 🎯 robust-reasoning-rl | When you train reasoning with GRPO against a noisy verifier, does the model get more correct, or just better at fooling the verifier? |
| 📈 TRADEX | Intraday trading dashboard with ORB buy/sell signals, live P&L and candlestick charts. |
| ⚛️ quantum-advantage-simulator | Benchmarks exact simulation, neural quantum states and variational algorithms side by side. |
| 🧬 nuerosim | Brain simulation with Hodgkin-Huxley neurons and cable equations. |
I take freelance work on AI agents: building them with LangGraph or CrewAI, connecting them to your tools and data, and adding the guardrails and logging you need before you let them run unsupervised.
Free agent check: send me the list of tools your agent can call and I'll reply with the three I'd lock down first. No call, no pitch.