We rearchitected our Browser Use agent to support users who do not want to, or cannot, run it entirely locally. Along the way we threw out the master plan, added a repository of skills, and started building toward something larger: a market where agents bid on tasks.
Bring your own endpoint
Running AI models locally requires a level of hardware and software that is not common. Mass adoption dictates that we allow someone to bring their own OpenAI-compatible endpoint and API key and start using it.
The repo is fully open source under an Apache 2.0 license.
The design has been radically simplified
Instead of making a full multistep plan and then executing the steps one by one, we now rely on the orchestrator to decide only the immediate next step. Then act, observe, and determine whether the step achieved the required result. Rinse and repeat. If you get caught in a doom loop, take a step back and do a strategic review.
But gone is the master plan at the start.
The reason is that, often just like in life, you make a plan and the actual reality turns out to be significantly different. What is needed instead is a general sense of what it takes to progress toward the objective, plus the immediate next step. You take the step, feel your way through, and keep progressing forward.
This lets the agent deal with fluid environments that may not match expectations. A master plan sets you up for potential failure and for time and token waste, because it does not allow for feedback as often as feedback is actually needed.
We have noticed that this harness is far more robust and delivers better performance.
Skills: the paths you have already trodden
Not everything is unknown. You do not need to feel your way after you have walked the same path many times.
This is where skills come in. We have developed an extensible repository of skills, which are sets of instructions for performing common actions: use Google Sheets, use Docs, make posts, collect data from social media, and so on. We ship some built-in ones and let you write your own.
You can demonstrate an action you perform often, and it gets added as a skill. This is akin to building a neural connection. Next time you don't have to figure things out; you can take the shortcut.
When posed with a task, the harness checks whether there is a skill it can rely on. Even then, the per-step feedback mechanism remains. A skill just means the agent has a better idea of how to do things.
We will also be allowing users to share skills, which means that over time, and as more users adopt it, the harness gets more and more powerful.
Traces that make the model better
We also have the option to collect PII-scrubbed data on successful traces, which will let us post-train our models. So not only does the system become more efficient as skills accumulate across users, the model itself gets better and better as we continuously post-train it on the RL data we generate.
That is a double win.
Always-on agents, and a digital human on your machine
We will be incorporating always-on, heartbeat-based agents, similar to OpenClaw and Hermes. The agents can then take on much longer-range tasks, work in the background, stop waiting on user input at every turn, and update you on their progress of their own accord.
This effectively creates a digital human on your machine: one with full capability over your browser, and eventually over your entire machine via computer use.
The task market
Which brings up the next interesting possibility.
These digital humans can participate in a task market. Koretex nodes move up from being pure inference providers to being independent operators that can perform an entire task for you. You would have layers of abstraction: you give a request of any complexity, and in the background it is bid for and pushed to the digital human with the most superior bid, who can then disassemble it into smaller tasks and distribute them further.
This will eventually be the future of work.