A task arrives with a harness — the tools it can call, the skills it can reuse, the specialist agents it can delegate to. A lead agent turns that into an orchestration, which acts in the environment; feedback returns to the lead agent and its persistent state. Nothing here holds still: the harness keeps changing, the system rewrites itself to match, and every version of the pair is superseded by the next. Run that far enough and the whole of it — every generation, and everything it kept — is one unit of the next thing being built.
The map that framed our whole agenda — from standalone LLMs, to single-agent systems, to multi-agent systems. The must-read starting point that everything below builds on.
The first to make MAS design automatic — an inference-time, self-refinement framework that builds multi-agent systems with zero supervision. The seed idea the rest of the line scales up.
Holistic, training-time orchestration via function-calling RL — composing an entire MAS at each step, not piece by piece — paired with MASBench, a controlled study of when multi-agent beats single-agent. Beats GPT-5 and Claude-Sonnet-4.5 by up to 23% across 5 benchmarks, at a 10× efficiency gain. It's the trunk the three directions below branch from.
Building on the orchestration line — skill-based agent routing, an offline approach that outperforms SOTA RL orchestrators by 22.5% at 700× lower cost.
Can we judge an orchestration without running the whole system? Reward modeling and process verification at the orchestration level.
When does multi-agent actually beat single-agent? A cost-controlled study plus benchmarks that stress-test the multi-agent advantage.
Everything above assumes the harness holds still. It doesn't. Can your agents keep pace with an evolving harness? — a benchmark that isolates the simplest slice of that problem, pure growth: tools, skills, and specialist agents accumulate around a fixed model, stage by stage, nothing is retired, and outer harness evolution is separated from the system's own inner adaptation. 17 harness streams, 802 tasks, 520 tools, 42 skills, 62 agents. Expansion alone costs frontier agents work they had already solved: harness-induced forgetting of up to 46.4% on the agent axis.
Where the whole series becomes something you can run. Describe a problem and Orchestra designs a multi-agent plan, executes it, and lets you refine it in chat — across math, search, deep research, and enterprise ops.
A lead agent configures the environment, orchestrates sub-agents, receives feedback, and updates persistent state. The hard parts that remain: Policy (a huge action space), Verification (judging a plan without a full rollout), and Environment — where EvoHarnessBench now gives us a measuring stick, and the open problem becomes bi-level adaptation: learning inside a stage in a way that still holds up once the harness moves again.