Turning a 26B open model into an auditable science research agent
- Source
- x.com
- Date
The @lokahq team post-trained our 26B Trinity Mini into an agentic harness specialized for scientific work. Working with @PrimeIntellect and @AWS, Loka taught Trinity Mini to execute multi-step research workflows, coordinate tools, query databases, and log an auditable trail at every step. Here is a look inside the harness. 🧵
From raw sequence to 3D structure. Specialized tools pull sequence data, run folding routines, and output structural visual insights in real time.
Unstructured chat transcripts fail in real scientific work. The harness was built to generate deterministic, structured artifacts instead. Commands like /plan or /report trigger fixed research workflows. An internal critic agent then steps in to stress-test the model's hypotheses against retrieved evidence.
It handles chemical structures with the same precision. Researchers can view and analyze 2D molecular properties right inside the interface without jumping between external applications or losing context.
Context
The Loka team post-trained Arcee AI's open 26-billion-parameter Trinity Mini model into a built for scientific research, working with PrimeIntellect and AWS according to Arcee's own account. The stated design choice is that the harness produces fixed, structured outputs, triggered by commands like /plan or /report, instead of an open-ended chat transcript, because Arcee and Loka say unstructured transcripts are unreliable for real scientific work. A separate internal critic is then used to check the model's conclusions against retrieved evidence before they're finalized.
Arcee describes the harness resolving ambiguous scientific questions by querying primary databases such as PubMed, UniProt and directly and its answers in what it retrieves, and it describes tool-driven workflows covering protein structure prediction and molecular property analysis. The thread illustrates these with videos that are not available for direct review here, so this account reflects what Arcee and Loka reported rather than an independently observed demo.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- generative engine optimization — Making a site the source an AI assistant cites when answering, rather than a link in a list of ten blue results.
- grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
Checking sign-in…
Loading comments…