Recently I have been thinking a lot about background agents. My primary curiosity asks how do we make background agents maximally useful in dynamic scenarios?

There are two strains of thinking: full autonomy and human-in-the-loop.

Full Autonomy

Zapier and n8n are leading the way for the construction of agents, which are just sets of actions with an AI navigating a decision tree. We have seen full autonomy work best for simple tasks with a small set of required actions.

Here is how I envision a full autonomy model working:

1. Use a smart thinking model
2. Construct a tightly constructed, clear workflow.
3. Allow for a wide margin of error

The issue with full autonomy is the high failure rate and lack of accountability when things go wrong. When agents perform tasks with many forking steps and complex failure conditions, they break down.

Human-in-the-loop

Cursor is the gold standard for human-in-the-loop background agents, which don’t act until prompted and periodically ask for confirmation before executing a task. I’ve experimented with Yolo mode in Cursor, which gives the agent more autonomy to fly around and make changes.

But all Cursor power users know best practices when using agents:

1. Only give the agent tasks that a junior engineer could figure out
2. Don’t confirm the change unless you understand what the agent has done

Context

I’ve been playing with lots of simple agentic consumer ideas — smart Pomodoro, deep work insight engines, and, most recently, web task completion tools. They all center around the consumer, and rely on deep understanding of who the consumer is, their daily trends, and how they solve problems.

Context is the new holy grail for building better agents. We already have data collation techniques, like ASTs and RAG, but the problem is the amount of high-quality context we can garner.

Of course, privacy is the main concern here. Users will need to retain fine-grained control over their data and transparency around what information gets sent to the model. There needs to be innovation to allow agents to flourish with maximized context but transparency for user protections.