Radient is a Toronto-based AI research and development company building products and solutions that lower the cost and barriers to access to business productivity assistance for small businesses and sole proprietors. With AI rapidly becoming a table-stakes requirement for business operations, there emerges a real existential risk of extreme imbalances if the technology is not evenly distributed to all.

Radient’s first product release was Radient Automatic: a model routing solution for AI agents. It integrates seamlessly into any standard agent workflow to automatically reduce costs. It assesses the domain and difficulty of each step of an agent’s work path and picks the right model for the job, switching in real-time based on difficulty and domain. By leveraging Inception’s fast and cost-effective diffusion Mercury model, Radient is able to power key components of the routing system that are responsible for predicting agentic task type and difficulty at every step.

“We cut routing and classification overheads to sub-second latencies even on complex agent traces which allowed us to build tools that make assistants feel fast while keeping the bills low for our customers.”
- Damian Tran, Founder, Radient

Radient Automatic is currently used in Local Operator: their open-source, multi-agent assistant platform with a growing community which integrates automatic selection by default so that end-users get the right model per step without needing to specify one. This is important for less technical users of agentic platforms who may not know which model is the best to tackle different parts of their request.

The Challenge: Latency at the Router

Agentic traces often branch and extend quickly for even relatively simple tasks: retrieval, quick summarization, structured extraction, code actions, and long-form drafting. Even within one request, an agent may need to do research, consolidate information, and write a final report for the user. Each of these parts of the task have different models that are best to tackle them.

If the router is slow to switch between models, overall UX suffers even if the chosen model is fast, and this slowness compounds with the length of the agent journey. Radient needed a sub-second routing signal that stays accurate across heterogeneous tasks that would work well within the rapidly updating machinery of the automatic routing system.

The Solution: Mercury as a Fast, Parallel Decision-Maker

Radient incorporated Inception Mercury as one of the components within the multi-model classification brain inside their Automatic solution:

1. Complexity assessment. For every agent step, Mercury generates structured metadata to classify the nature (e.g., “RAG-style lookup vs. creative drafting vs. structured parse vs. tool/action”) and difficulty, plus a tight rationale.

2. Routing decision. Radient Automatic has a table of internally benchmarked models that constantly updates with ongoing testing and customer feedback. The system maps Mercury’s assessment to a model policy (e.g., lightweight fast chat for summaries; heavier reasoning for complex synthesis; specialized models for code or vision).

3. Step-wise continuity. The same process repeats step-by-step so routing adapts as the agent’s plan evolves.

Mercury’s diffusion architecture generates text in parallel passes instead of strictly token-by-token, which is precisely what Radient needed for fast structured generations. The Mercury generates text at over a thousand tokens per second and runs 5 to 10× faster than comparable light autoregressive models while matching their quality. This was ideal for Radient’s “router in the loop” use.

Outcomes for SMBs and Solopreneurs

• Speed that feels instant. Sub-second routing means that users don’t notice significant slowdowns compared to using single models on long horizon tasks. That keeps users in the flow and makes model routing make sense for the user experience.

• Lower costs automatically. Radient Automatic routes simple steps to inexpensive models and reserves heavier models only when needed. Radient’s internal testing showed that this resulted in up to 70% cost savings on coding, research, and administrative tasks as compared to using a heavy model the entire way.

• Scales with your work. As tasks get more complex, per-step classification scales the model selections in complexity without requiring the developers of agentic platforms to re-architect their flows or rewrite prompts.

Why Mercury

Mercury made sense within the Radient Automatic flow for making a lot **of small, correct decisions fast on unstructured text. Diffusion-style generation gives Mercury a parallel refinement path that’s well-matched to short, structured outputs like “task type, difficulty, rationale.” In practice that means the fastest path to a schema-friendly answer that plays well with other systems.

Radient evaluated lightweight autoregressive options, but they tended to carry higher latency overheads on mid-sized structured outputs. Mercury’s design gives predictable sub-second behavior even under bursty agent workloads. The result is a cleaner separation of concerns: Mercury excels at the micro-planning loop so that Radient Automatic can map that decision to the right specialist or generalist model.

Takeaway

For SMBs and sole proprietors, the only metric that really matters is time-to-useful-answer at a price that feels fair. Embedding Mercury inside Radient Automatic lets Local Operator assistants stay responsive on long, branching tasks because the classification step is effectively invisible. These decisions happen under a second, so users experience steady progress rather than pauses between model switches. That speed, combined with routing that reserves heavy models only when the task demands it, is what turns “AI” from a line item into a daily advantage: faster inbox triage, cleaner research summaries, fewer hours lost to admin.

Mercury is a key piece of that puzzle: it keeps the router decisive and cheap while Radient’s policy layer and billing unify everything behind a single, predictable flow. The advantage for users is frontier-grade help, delivered at small-business scale and cost, without asking users to learn the difference between twenty models just to get work done.