Insights
Founder and builder perspectives on AI tools, thinking patterns, and the new way of working
Showing 121-132 of 485
When to Fork an Agent vs Spawn a Fresh One
Context-inheriting forks and stateless fresh subagents solve opposite problems. Picking wrong either wastes tokens or loses critical context.
The 15x Token Tax of Parallel Subagents (and When It's Worth Paying)
Parallel subagents cut research time up to 90% and cost roughly 15x the tokens of a single chat turn. That ratio is a budgeting decision, not a red flag.
Orchestrator vs Swarm: Two Ways to Split Agent Work
Hub-and-spoke and peer-to-peer aren't rivals fighting for the same job — they solve different shapes of problem. Here's how to tell which one you have.
Why the Lead Agent Pattern Won in 2026
Flat agent swarms looked elegant on a whiteboard. In production, hierarchy won because someone has to own the failure.
A Context Engineering Checklist for Your Next Agent
Nine questions to run through before shipping an agent, pulled from the mechanisms that actually cause most agent failures — attention budget, tool sprawl, memory, and compaction.
The Case for Structured Memory Over Bigger Context Windows
A bigger context window doesn't fix the attention problem — it just gives you more room to spend badly. Structured, retrievable memory does what raw scale can't.
Writing Skills That Load Context on Demand, Not by Default
The best agent skill libraries look small from the model's perspective even when they're huge — because almost nothing is loaded until the task actually needs it.
Sub-Agent Isolation as a Context Engineering Pattern
Spinning off a sub-agent isn't just a way to parallelize work — it's a way to keep a noisy investigation out of the main agent's context window entirely.
Context Compaction: What to Keep When the Window Fills Up
Every long-running agent eventually has to summarize or evict something. The teams that get this right decide deliberately what survives compression — not by default.
The 20-Tool Ceiling: Why More Tools Make Agents Dumber
OpenAI's own guidance caps recommended agent toolsets under 20, with accuracy degrading past 10. Here's the mechanism behind that ceiling, and how to prune a toolset that's already grown past it.
RAG, Skills, and Tools: Three Ways to Get Information Into Context
Retrieval, packaged skills, and tool calls all solve the same problem — getting the right information in front of the model at the right moment — but they trade off very differently.
Why Most Agent Failures Are Context Failures, Not Model Failures
When an agent gives a wrong answer, the reflex is to blame the model. In 2026 the more accurate diagnosis, most of the time, is that it never saw the right information.