Gopichand Kotana

The current AI tools and products are very, very good. They got particularly good when people realized that LLMs are good thinking machines and used them mainly for reasoning.

Encapsulating an LLM as an agent and supplementing it with everything that’s not reasoning, unlocked AI across many domains. LLMs are no longer thought of as chatbots; they are the thinking component of software for software that needs reasoning.

The AI landscape is evolving rapidly, and most new AI products that emerged in the last few months seem to converge around three areas: knowledge augmentation, Context Engineering, and reusable skills/workflows (giving AI access to known/derived workflows from the existing system). The boundary between all these ideas is fluid. MCPs briefly got super popular, but now they’re being wrapped into sub-agents. Agents are becoming the abstraction layer that composes all these ideas.

The industry also seems increasingly comfortable with the idea of the agentic loop. At its core, the loop is simple: an LLM does the reasoning, while surrounding systems provide tools, knowledge, memory, constraints, and evaluation.

Most of the engineering challenge now is about maximizing the quality of that loop, giving the model the right context, grounding it effectively, monitoring execution, and deciding when the output is good enough to trust. And we are all trying to get the most out of this loop.

I think the next few months/years will be about agent orchestration and communication. We have to figure out how agents communicate, delegate, verify, recover from failure, and stay aligned with their given objective. It is not very different from microservices, but unlike traditional services, the agents have non-deterministic contracts, so we need to keep them grounded and ensure they don’t stray too far away from their tasks. Probabilistic contracts also mean that a fully decentralized system with peer-to-peer communication is hard to build while making sure that the output of all sub-agents aligns with the final output. Centralized coordinators seem to be the way forward for now.

The next few blogs will be more technical, covering aspects of Agent coordination.