#67 Prateek Karnal - Orchestrating the Agent Army

October 5, 2026

Ben

46:00

mins

Listen Now

46:00

mins

Prateek built Agent Orchestrator (AO) while working on AI infrastructure at Composio, where agents were already running at scale. When he tried the same on his own laptop, more than three agents and he lost the thread. The personal fix went viral, passed 12,000 GitHub stars, and turned into a company.

Most developers lose track somewhere around three parallel coding agents. Prateek Karnal runs ten to twenty, and he built the tool that makes that bearable.

Prateek built Agent Orchestrator (AO) while working on AI infrastructure at Composio, where agents were already running at scale to build the company's integrations. When he tried the same on his own laptop in February 2026, more than three agents and he lost the thread. The personal fix went viral at launch, passed 12,000 GitHub stars, and turned into a company. Ben and Prateek get into what running a fleet of agents involves once the screenshots stop: where the human belongs, what the interface has to show, and why the hard part is shifting from doing work to finding it.

About this episode

Prateek runs about ten agents locally at any given moment, a mix of Claude Code, Codex and Gemini CLI, picked per task, and he puts his own productivity gain at two to three times, well short of linear, because intent and output still need a human to verify them. The parts he wants automated are the mind-numbing ones: figuring out which agent is doing what, switching tabs, and copy-pasting CI failures into terminals, which he suspects most engineers still do by hand. His thesis for AO is an optimized human in the loop. Full autonomy, in his view, is off the table for now.

The bigger argument starts with the most common pushback at launch: "I only do two or three tasks a week, why would I need this?" Prateek's answer is to use the agents to find more work. Production errors, GitHub issues, customer signals: a software company produces more information than one person can process, and an orchestrator agent can turn it into a plan, spawn workers, and hand you green PRs to verify. Ben connects this to the cubicle problem. In most large organizations people wait to be prompted by their manager, and Prateek's take is that this is the mindset most likely to get replaced.

"The biggest question you need answered when you're running a fleet of agents is: what should I look at right now?"

AO's interface took three iterations: no UI at all, then a list of agents, then a Kanban board after a product manager friend at Composio suggested one. Each card is a piece of work with one agent on it, showing the harness, token spend, activity and CI status. The columns were the hard part. No human moves the cards, so position has to be derived from what the agent has achieved: working, PR out and iterating on CI and review comments, green PR waiting for a human, approved. Agentic testing, where the worker deploys and posts a screenshot of what it verified, is next. Ben, who spends a lot of his working life in front of Kanban boards with humans on them, finds this both familiar and strange. In a normal system different people pull a ticket through different stages. Here one agent drags it across the whole board.

A founder of an AI security company in Texas forked AO months ago, built the missing features himself, deployed it on a cloud machine, wired it to Slack and ClickHouse, and now runs around 500 agents that his employees spawn by tagging a bot. Prateek took that as validation of the enterprise thesis before a single customer paid. Ben has his own cautionary tale: a tender agent with a limited credit card that logged into Stellar Work's insurance account and bought a bigger policy for about a thousand euros, because a public tender required it. Stellar Work did not win the tender.

"Did my robot just spend a thousand bucks to do its job?"

AO is built for how agentic engineering will look in six months: mobile and cloud, chat UIs for more than 25 harnesses, and an orchestrator with long-horizon memory that merges low-risk PRs on its own when the signals are strong. Prateek's analogy is a self-driving car that handles the freeway and hands back the wheel at a tricky turn.

Key takeaways

Productivity does not scale with agent count. Prateek's honest estimate for himself is two to three times, because verifying intent and output stays human work.

Once you can parallelize, the bottleneck moves from doing work to finding it. An orchestrator agent reading production errors and GitHub issues can help decide which ten tasks matter.

A Kanban board works for agents because the columns can be derived from what the agent has achieved: PR out, CI green, review passed, approved. The cards move themselves.

Two signals deserve your attention first: an agent blocked on a question, and a green PR waiting for a human. An agent grinding through CI failures is doing fine without you.

About the guest

Prateek Karnal is the creator of Agent Orchestrator and founder of Untrivial, the company behind it. Before going full time on AO he built AI infrastructure at Composio and led its integrations layer, which is where he first ran agents at scale. AO is free and open source under Apache 2.0 and supports more than 25 coding-agent harnesses. He wrote up how the first version was built, 30 parallel agents in eight days, in Open-Sourcing Agent Orchestrator.