Kojo Osei

 undo All posts
Subscribe

Your agent needs an inbox

• 2 min read

Conductor is a coding IDE that lets you run a bunch of coding agents in parallel. If you're used to tab autocomplete or terminal-based agents like Claude Code, Conductor's approach feels fundamentally different.

Here's how it works: agents write code for specific tasks in isolated workspaces. You break your project into chunks, dispatch them to agents who work simultaneously, then review and integrate what they produce. This works best when you have independent tasks that can be clearly defined upfront.

Using Conductor gives you a preview of what knowledge work might look like when we routinely orchestrate multiple agents. Which raises a question: what will the dominant interface for dispatching and reviewing agent work actually look like?

I think it'll look a lot like an agent inbox with the following pieces:

  • Specs layer: A place to write clear, reusable instructions for each task
  • Agent dispatch: A simple way to pick the right specialist agent (e.g. for refactoring, research, etc.) for each job and perhaps set a compute budget.
  • Status monitoring: Real-time visibility into what your agents are doing. Are they stuck? Making progress? Waiting for input?
  • Embedded evals: Built-in tools to review an agent’s work, give feedback, and integrate it.

This "dispatch, monitor, evaluate, integrate" pattern extends far beyond coding. Picture a deep research agent building an investment memo. You dispatch it to analyze SEC filings, synthesize analyst reports, and model market trends. You monitor its progress as it parses documents, getting alerts if it hits contradictory data.

Today, many teams try to retrofit agent orchestration into tools like Slack. This feels intuitive — it's where we already work. But it creates friction. Slack is designed for conversational flow, not structured task management. Its primary unit is a message in a chronological stream. You cannot see, at a glance, a dashboard of which agents are active, what phase of their task they’re in, or if they’re blocked. 

A true agent inbox, in contrast to today’s retrofit, gives you a place to orchestrate intelligence effectively.

In a world where you could spin up 100s of agents for 100s of analyses, the critical questions change. How do you prioritize which tasks get the fastest compute? When is it worth $10 in API calls versus five minutes of human review? Which results deserve trust, and when should a human step in? An agent inbox becomes the control plane for managing these new workflows.