Microsoft bets agent work goes multi-model, and the harness decides which model earns each job

What Microsoft is building

Microsoft is introducing a Copilot harness to facilitate multi-model agentic work, and the company says the harness is designed to solve for performance, quality and cost in AI model integration. That framing matters because it treats the model as only one part of the product. The harness is the part that has to choose, sequence and judge the work across models. m:…9302

Microsoft describes building a system that works seamlessly across multiple models and surfaces as incredibly hard. That admission is the honest core of the announcement. Anyone can call two models from one script. Making the combination feel like one reliable system, with consistent quality and predictable cost, is the work that usually fails quietly. m:…9302

Tina Schuchman and Vladimir Fedorov assert that competitiveness requires a multi-model approach. Put plainly, the argument is that no single model wins every job, so the team that routes well beats the team that picks one model and hopes. The implication for buyers is that model selection stops being a one-time purchase decision and becomes an operating discipline. m:…9302

Why handoffs beat loyalty to one model

Research on model collaboration gives the multi-model argument its sharpest form. AgentBaton found that capable-to-efficient model handoffs at suitable stages of task progress preserve or improve accuracy while substantially reducing cost. The direction of the handoff is doing the work here: start strong, then hand off once the shape of the answer is settled. m:…1714

The same study found the reverse order does not work as well. Efficient-to-capable handoffs generally underperform capable-only execution with limited or no cost reduction. That asymmetry is a useful rule of thumb for anyone designing an agent pipeline. Cheap first is not the same as cheap overall if the weaker model sends the stronger one down the wrong path. m:…1714

Across its tests, AgentBaton reduced API costs by 65.9% on average while largely preserving task accuracy. A number that large should still be read as a research result in a preprint, not a promise about any particular workload. But it sets the stakes: if routing is done well, the savings are large enough to change which products are viable, not just to trim a bill. m:…1714

Diversity without swapping models

Multi-model work is not the only way to get diverse reasoning. Agent-conditioned routing produces diverse reasoning paths within one frozen model, without user-assigned roles, prescribed methods, fine-tuning or model copies. On one measure, it increased the number of problems yielding at least two distinct primary methods by approximately 2x on GLM 5.3 and 1.67x on GPT-5.6 Sol compared with standard temperature sampling. m:…8904

The cost of that diversity looks small in the preprint: about 160 KiB of mask state, with latency and peak VRAM usage within 3% of batched temperature sampling. That matters because it separates two different bets. One bet is that better answers come from more models. The other is that better answers come from more ways of thinking inside one model. A serious harness probably needs both, and should be judged on which one it uses when. m:…8904

The long-horizon view comes from HaMo, a framework for long-horizon continual learning in web agents that combines non-parametric experience accumulation with parametric model adaptation. At the harness level it distills trajectories into compact experience cards, stored in persistent non-parametric memory for immediate retrieval. At the model level it updates a parameter-efficient adapter using selected trajectories and historical replay. m:…9564

HaMo outperformed harness-only adaptation in within-domain and cross-domain shifts, and it outperformed model-only adaptation in the same settings, while improving retention of previously acquired capabilities and adaptation to new tasks. The lesson is not that harnesses replace models or vice versa. It is that the memory of past work and the weights of the model have to improve together, or each one caps the other. m:…9564

What to watch

The open question for the Copilot harness is whether its routing judgments hold up on real workloads, where quality is harder to score than in a benchmark and cost depends on traffic patterns nobody publishes in advance. Microsoft has named the three things the harness must solve, performance, quality and cost, which is also a useful checklist for anyone evaluating it. A harness that is fast and cheap but quietly worse is not a bargain. m:…9302

The research points to a second test: does the system know when not to hand off, and when to stay with the capable model? AgentBaton suggests the wrong direction costs accuracy without saving money, and HaMo suggests experience has to be stored in a form the system can actually reuse. Teams adopting multi-model agents should therefore ask for handoff logs, not just model lists. If you cannot see why a job went to a particular model, you cannot fix it when it goes wrong. m:…1714 m:…9564