Narrow Agents
Narrow Agents
Sandeep Swadia's thesis that the agents that win — in both the enterprise and the startup market — are obsessively narrow, not broadly capable. From You're Not Behind (Yet) Learn AI Agents (theMITmonk).
"Most companies want AI everywhere. The ones actually winning are obsessively narrow. If clarity is the bottleneck, then the opportunity is not broad intelligence, it's narrow ownership."
The construction-software anchor
Customer conference for a construction-software company. Product lead demos a beta agent that solves one specific field-data-collection workflow for one specific kind of customer in one specific situation. The demo glitches a bit. Doesn't matter:
"When he showed the QR code at the end, every hand in the conference room went up with their phones. Everyone took a picture because it solved a very specific but very real pain they all had been living with for decades."
The lesson: depth of fit to a known-painful workflow beats breadth.
The founder thesis
"For every software company that exists today, there will be an agent company trying to dethrone it. The winners won't build the broadest agents first. They'll build the one that understands one workflow, one market, and one kind of user pain better than everyone else."
Sandeep's test for spotting the opportunity:
"Find a highly specific task people hate doing, but they have to do it repeatedly. That's where the money is."
Three conditions: highly specific, hated, repetitive. The first two surface real pain; the third makes the unit economics work.
How it composes with existing vault frames
- Build vs Buy (Agents) (Praveen Akkiraju) — Sandeep's narrow-agent thesis aligns with the back-end build side: domain-specific agents that no vendor can pre-build. Adds the founder/product-strategy lens to what the existing page covers as enterprise architecture.
- Bounded vs Unbounded Tasks — narrow agents are nearly always bounded (one workflow, one pain). The narrow-ownership thesis is partly a restatement of the "start with bounded" rule with a stronger commercial framing: don't just start narrow because it's safer, stay narrow because that's where the value compounds.
- ARR Framework + GPS Check (for Agents) — the agents that pass GPS are the ones where one person/team can fully articulate goal/proof/steps. By definition that's a narrow scope. Sandeep's GPS-then-narrow chain is internally consistent.
- Code Is Free — when implementation is cheap, the moat moves to what specific pain you understand better than anyone else. Narrow-ownership is that moat in agent form.
Implied warning for broad-platform players
The video implicitly argues against the AI-strategy pattern of "deploy AI everywhere across all functions." That approach maxes out token-spend (see Token Maxing) without ever converging on a workflow the agent understands deeply enough to be reliable. The 6,000-agent Cvent number (Shadow AI) is a positive case of intentional sprawl as learning; the Sandeep-warned anti-pattern is unintentional sprawl as substitute for focus.
Practical takeaway for this vault's user
- Brand fodder candidate — "Most companies want AI everywhere. The ones actually winning are obsessively narrow" is a near-ready Medium post hook. Pairs cleanly with CIO Agenda 2026 (CXOTalk)'s 88%/<6% gap as the data point.
- Org-coaching artifact — for any team-member proposing a "general AI assistant for X function" agent, the question becomes "what's the one workflow within X that's most painful and repetitive?" — and start there.