2025 — ongoing
seven months, one promotion, a million learnings: moving across the country to build ai agents
i packed up, left everything familiar, and landed in new york city to work on something i believed in. seven months later i had a promotion, a fully different understanding of what it means to build ai, and more questions than when i started — which is usually the sign you are in the right place.
what we are building
amigo is an ai agent platform. not a chatbot wrapper, not a prompt template — an actual agent layer that reasons, plans, and executes across tools and contexts with minimal hand-holding. the product is built for teams that need ai to do real work, not just assist with it.
my role sits at the intersection of agent engineering and product. i help define how agents should behave, what they should have access to, and how to build the feedback loops that actually make them get better over time.
what i have learned about agents
the gap between a demo and a reliable agent is enormous. demos are easy — one happy path, controlled inputs, a compelling output. production agents are a different problem entirely. they fail at edges you did not anticipate, drift when context gets long, and hallucinate with confidence when uncertain.
the agents that actually work share three things: tightly scoped objectives, clean tool boundaries, and observability that lets you catch failures before users do. building agents is less about prompting and more about systems design.
the tooling problem
the hardest part of building agents is not the model — models are getting better fast and cheaply. the hard part is tooling. an agent without tools is just a text predictor. an agent with the right tools at the right time is a collaborator.
most tooling infrastructure is static: define your tools, register them, hope they cover the use case. the more interesting question — the one i have been thinking about most — is what happens when agents can build and register their own tools dynamically, on demand, for the problem in front of them. that is the unlock that makes agentic workflows truly non-deterministic and uncapped.
why this moment matters
we are at an early and consequential point in the agentic era. the infrastructure choices being made right now — how tools are defined, how agents communicate, how trust and scope are managed — will set the defaults for what ai systems look like for a long time. working at amigo means being inside those decisions while they are still open.
that is not a small thing.