March 26, 2026
Funny how we standardized on the word "harness" to describe the environment in which agents are executed and the human/agent interface associated with it. The word harness suggests a handle for humans to box and control agents to exploit their power. But when intelligence sits on both ends of the interface, it is no longer a one-way handle but rather a bi-directional medium. Maybe the cold truth is that the harness is probably as much for us as it is for agents[0]!
Purposefully or not, a harness is also inherently a single-player concept. 3 years ago I wrote[1]:
ChatGPT is the Pong of LLMs. Next we'll inevitably go through a trough of disillusionment… … but imagine, one day we'll get the DOOM, Civ, Red Alert, and Counter Strike of LLMs🤩 Let alone multiplayer modes🔥
Continuing with that analogy… With Dust, we're building something like the level editor for Pong… whatever that means🤔 Now that gets me thinking about multiplayer mode 😬
I would concede that with the emergence of agents and improvement of their harnesses, including sandboxes and access to a plethora of tools, we've moved from Pong to DOOM; or even, let's be bold, Civ and Red Alert in the still rare cases where many agents are used in parallel. But we're still very far away from working multiplayer modes.
Most, if not all, human-agent interactions today are still single-player. For the past three years, despite effectively pushing collaborative features in Dust (most notably what we would describe today as the ability for multiple humans to interact with agents through a shared harness), our users (and us) still have a tendency to avoid working with agents collaboratively most of the time. We revert to the single-player mode, nudged by an invisible force.
Why is that? Two hypotheses:
(i) Interacting with an agent feels very personal due to its messiness. Because agents and their harnesses are still very much imperfect, getting to a specific outcome often requires a lot of gnarly impure pedestrian steering. Put differently, you don't want to see how the sausage is made, and we all have our own way of making it.
(ii) The typical time-horizon of the tasks we use agents for is still pretty short. Collaboration does not happen in the task, it happens outside of the task (On GitHub for code, on Salesforce/Hubspot for sales, on Slack, Notion, Google Drive and many other platforms for everything else).
Yet, the time-horizon of the tasks that agents can tackle is rapidly increasing[2]. It seems unlikely that steering agents working on weeks or months-long human equivalent tasks won't require collaboration from multiple humans within an organization. The force vector of (ii) is about to switch direction pushing us towards multiplayer collaboration instead.
When that happens (i) and (ii) become at odds and will require new human-agent interfaces to resolve the tension. Collaboration will happen within the task, aka the agentic loop, instead of outside of it. We'll need aside interactions operating on shared context and state. We'll need to provide agents with ways to interface not with one human but the whole organization so that they can ask questions, assign blocking sub-tasks, and more broadly orchestrate work, parallelizing when possible, blocking when needed. We'll need ways for teams of humans to collaborate on introspecting work and steering agents as they make progress. The harness will go from a human/agent harness to a company/agents harness, an operating system for agents/agents and agents/humans collaboration.
Shaping the right interfaces for this collaboration to happen is the next frontier, and this frontier feels excitingly unscratched as of today.
[0] The word harness becomes all of the
sudden a bit more uneasing when seen through that lens.
[1]
https://x.com/spolu/status/1625773663290068995
[2]
https://metr.org/