Humans and Agents

Right at the dawn of the agentic AI hailstorm that started in 2024, I had a chance to be part of something paradoxical - reimagining the human-in-the-loop layer for business agents, whose primary premise was to reduce, or eliminate humans in the loop

Or so I thought, and so a lot of the world thought

But peak business productivity isn't eliminating humans, it's eliminating human monotony. Empowering humans to do the things only they can do - dream, define and direct. Everything else - can be 'automated', or 'agentified'

And thus, the right business system doesn't eliminate humans - it makes way for human entry to be as seamless and objective as possible. It reduces human limitations - cognitive ability and subjective fatigue, while ensuring 'real world sense', control and critical thinking continue to be fully human steps.

And that's what I worked to build - a human in the loop system that chose to 'blend' rather than be a sore standout. A human intervention process that feels as natural to a complex, agentified business process, as an agent.

And if you solve for this experience, only can you solve for businesses.

So what did I build? And how?

Did I recreate the entire product? The product that had once been a rules-driven, structured and deterministic human layer, into a cog that oiled the agentic machine? I wanted to, but the world works differently. For all its hype, agentic AI is still to prove its business effectiveness, and businesses are yet to trust their mission critical systems to them. Until that fortuitous dawn, our product had to cater to the rule based system that continued to serve the core of our business, while readying itself for the next shiny thing.

Here's how I thought about it -

1. I solve for the business,
2. I make users feel less overloaded and
3. I don't introduce a learning curve,

while keeping the engineering fundamentals streamlined -

1. my product scales effortlessly,
2. my COGS (infra and dev costs) are in check and
3. my architecture scales without refactoring - a human makes 10s of API calls - an agent might make millions. A human doesn't need audit of what they themselves did, an agent does

By keeping my fundamental rules clear, I crafted an experience - a chat driven, uniform, AI native human in the loop layer that was more than just another chatbot. It spoke, but it also acted - it was not just a cog in the agentic machine where humans were bit part players, it made humans the center of the cycle - one click was all it took, one voice recording was all it took, and MCP tools, efficient indexes across historical data, agent traces all cojoined to give the human the oversight that businesses sought.

This was a system that allowed business users to resolve 'escalations' - where an agent couldn't use its training and context to close a business operation, it had 'guardrails' - where agent tool calls that tampered with critical infra - Database calls, SAP warehouse updates were always human verified, and feedback and memory was ingrained - a human could teach an agent, similar to how you onboard a new employee, all while keeping the user interface streamlined, so users wouldn't wake up one day to see some random UI clogging up their human systems.