Rootconf 2026 Annual Conference

Abstract

Most demos show an agent completing one happy path. Real work is messier: long tasks, bad tool calls, context that overflows, and “it worked yesterday” failures.

In this talk I’ll walk through the harness I use when working with AI agents day to day: how I structure tools and memory, where I put human checkpoints, how I catch loops before they burn tokens, and what I measure so I know the agent is actually helping.

You’ll leave with a concrete checklist you can apply to coding agents or product agents (including agentic analytics / MCP-for-data workflows), plus a few failure modes I keep hitting so you don’t have to discover them live. No framework tour — decisions and tradeoffs from trying to make agents useful outside a notebook.

Outline

1. Hook — prompt-only agent that looped/burned tokens on a real task
2. What a harness is — tools, memory, retries, checkpoints, evals as one system
3. Tools & memory — tool vs context; short vs longer memory; one pattern against context rot
4. Human checkpoints — where to force pause (risky actions, spend, irreversible writes)
5. Catching loops early — repeat tool calls, no progress, cost spikes
6. What I measure — success rate, steps-to-done, intervention rate
7. Takeaway checklist — 5 stealable items for practitioners

Speaker bio

Vaibhav Arora is a Chandigarh-based builder working on AI agents and products that use them day to day (including JobAppAgent and SharedMoney / Glass Money). Ex-Snorkel. Focused on practical harnesses: tools, memory, checkpoints, and evals that keep agents useful outside demos. Speaks on AI agents and working with AI.

Audience and format

A talk — 30–40 mins. For Platform Engineering, SRE, DevOps, infrastructure, and engineering teams building or operating AI agents in production. Location: Chandigarh, India.