Saturday, March 14, 2026

Everyone's trying to give agents new 'skills' (like email or browser control) and get them running locally. But the real friction is orchestrating these local agents and their separate skills into *reliable, persistent workflows* without them losing context or needing constant re-prompting. Think about building a 'control panel' or an IDE plugin that lets you define and manage reusable 'agent playbooks' for specific local tasks, ensuring they *always* use the right skill at the right time and remember past interactions. You could start by creating a simple 'context manager' for local agents that automatically feeds them relevant files or browser tabs based on the current task, a bit like the 'local context folder' mentioned but without the upkeep pain.

Everyone's trying to get AI coding agents like Cursor and Replit to do more, but they constantly hit a wall because agents don't understand *why* certain architectural decisions were made, not just *what* the code does. You could ship a micro-tool that lets engineers quickly tag commit messages or PR descriptions with the 'why' behind major decisions, then serve that context directly to their AI coding assistants. This gives agents the 'institutional memory' they completely lack right now, making them exponentially more useful and less frustrating.

Everyone's hyped about AI agents getting giant 1M context windows, but they're still clumsy and expensive because they don't learn or manage that memory well. Instead of building another agent, make an 'agent brain optimizer' that sits *between* the agent and the AI, learning from past runs to automatically prune irrelevant context and inject *only* the crucial info for a task. You could ship an initial version as a local proxy or browser extension that logs agent interactions and suggests better prompts, helping builders cut costs and make their agents actually 'smarter' over time.

Everyone's getting AI to write code fast, but the dirty secret is those tools suck at generating real-world, end-to-end tests and making sense of old codebases for new hires. The timing is perfect to build an AI agent that specializes in ingesting an existing codebase, understanding its specific architecture, and then either auto-generating high-quality, user-scenario tests or creating dynamic onboarding guides for new developers. You could start by hooking into GitHub repos and using RAG (retrieval augmented generation — letting the AI look up specific documents) to feed it existing code and documentation, then focus on generating test cases that simulate user flows or context-specific explanations for new team members.

Every builder is racing to ship, but the moment you need secure user logins or reliable connections to multiple services (APIs), you're suddenly an infra engineer. What if you could build a super-simple wrapper around open-source 'patch' tools that just *adds* these features to your V0 or Replit project with a few clicks? The specific gap is making enterprise-grade identity (like 'Login with Google' for *any* app) and API management as easy as installing a browser extension. The first person to productize that for the low-code/no-code crowd will win.