# Agent reliability

> Making tool-using software understandable and dependable.

Canonical page: https://rajakarthik.com/work/agent-runtimes/

## Overview

When software can make decisions and call tools, the difficult questions often concern boundaries: what it can do, what it should ask, and how it should recover.

I am interested in safe tool interaction, explicit state, meaningful human control, and evaluations that capture real task outcomes.

## Topics I explore

- Tool selection and clear permissions
- Recovery from errors and interruptions
- Human review at meaningful decision points
- Testing behavior across longer interactions

Useful agents should be understandable when they succeed and when they fail.