Entangle

What it does

Matching researchers to company problems by hand means reading 1,047 researchers' publications, projects and patents against the challenges of 153 companies: about 7 million tokens in the EINSTEIN dataset. Entangle gives that work to a team of 20 Mozaik agents. Analysts turn each company challenge into a research need. Profilers read only the researchers shortlisted for it. A Matchmaker pairs them, or says "no good match" and why. Writers draft an industrial PhD topic or a joint project for each chosen pair, and a Funding Scout adds open EU calls and related EU projects. A Critic checks every claim against the dataset record it cites and sends unsupported claims back to be fixed, so each sentence in a proposal has a source. A person approves each pair before anyone is contacted. A company that is not in the dataset can also describe its problem in its own words and get researchers back.

How the agents run concurrently

Which agents: 20 agents in one Mozaik runtime: 4 Industry Analysts, 6 Research Profilers, 1 Matchmaker, 3 Proposal Writers, 2 Critics on different models, 1 Funding Scout, 2 Shortlist participants and 1 Orchestrator. Every role works at the same time. What they share: one blackboard (MatchState) with the needs, profiles, candidates, pairs, proposals, verdicts, tasks and budget, keyed by dataset id. Every fact on it is a claim with the source ids it cites. Agents never call each other. They write to the blackboard and announce each artifact on one event bus that every participant hears. How they coordinate: there are no stages. Each task is its own Mozaik loop with a fresh context, so one agent can run several tasks at once. The Orchestrator makes no model calls and reacts to events. Every company goes to an Analyst at the start. A need is shortlisted the moment it is posted, and its candidates go to the Profiler with the fewest open tasks (each researcher is profiled once). When a need's profiles are in, the Matchmaker gets it. When a company's need, profile and pair have all passed the Critic, a Writer starts on that pair while other companies are still being matched, and the Funding Scout works on the same pair beside the Writer. The Critic reviews each artifact as it lands and sends rejected claims back to their producer as a new task, so fixes run beside new work. Only the limits slow it down: one runner caps calls in flight and rate limits per provider, and gives each free slot to the work furthest down the pipeline, so a company that is almost done does not wait behind new ones. Measured in the reference run (8 companies, 696 model calls): 2 h 47 min of model calls in 49 min of wall time, 3.4 calls at once on average under a cap of 6 sessions, and up to 5 roles calling a model at the same moment. The first proposal started at 5:56, before most companies were matched. Every event is in trace.jsonl, and the Timeline shows one bar per model call.