Research project · Agent reliability · Token efficiency
AI Workforce
Can a team of specialist agents carry a multi-step goal end to end, reliably and at a predictable cost?
In progressAI Workforce is our testbed for multi-agent coordination. A goal is handed to a set of specialist agents - research, product, marketing, sales, customer success, analytics - which have to split the work, hand off context and finish without a human stepping in. It is where most of our reliability and efficiency questions get asked first.
What we're investigating
Coordination between specialists
How agents divide a goal, what context has to pass between them at each hand-off, and where coordination breaks down as the number of agents grows.
Degrading instead of dying
When one tool or agent fails partway through a long run, how the rest of the system can carry on with reduced capability rather than discarding work that has already cost money.
Cost per goal
A single request can trigger many times its own token count in hidden loops, retries and self-correction. We measure that multiplier and test a layered hierarchy - cache, deterministic code, targeted data fetches, small models, then frontier models - that only escalates when a cheaper layer cannot answer.
Approach
- 1Run real multi-step goals through the system, not single prompts.
- 2Log every agent action, tool call and hand-off as structured events.
- 3Count tokens per layer, so a saving can be traced to the layer that made it.
- 4Classify every failed run by cause before changing anything.
Where it stands
In active development as a research system. It is not available to use, and there is nothing to sign up for.
Working on the same problem?
We'd like to hear from researchers, engineers and teams who have run into it too.
Get in touch