Our Framework

Our Self-Improvement
Framework

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here holds still: the harness keeps changing, the system rewrites itself to match, and every version of the pair is superseded by the next.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.

01 Task + Harness
Each run begins with a task and the harness available to solve it — the tools it can call, the skills it can reuse, and the specialist agents it can delegate to.
02 Lead Agent
A lead (meta) agent receives the task. Its job is not to answer directly, but to decide how the work should be done.
03 Orchestration
The lead composes and coordinates specialized sub-agents — assigning roles, context, tools, and a flow between them.
04 Environment
Sub-agents act in an environment, calling tools and producing intermediate results that become a response.
05 Feedback
The environment returns feedback. This signal is how the system learns whether the orchestration actually worked.
06 Persistent State Across
Memory, learned skills, and routing preferences the system keeps for itself. It sits outside the loop because it belongs to neither level alone: the loop writes it, but it outlives every harness — generations end, this only grows.
07 Harness Evolution Outer
Tools, skills, and agents are added, revised, and retired as the harness moves to its next version. Each turnover ends one generation and starts the next — and they come faster as they go.
OUTER · HARNESS TOOLS SKILLS AGENTS V1 INNER · ADAPTATION Task + Harness Lead Agent Orchestration Environment Feedback ACROSS · PERSISTENT STATE
Scroll to our roadmap
2025 · SURVEYLLM Reasoning Survey 2025 · MAS POLICYMAS-Zero 2026 · MAS POLICY · ANALYSISMAS-Orchestra 2026 · MAS POLICYSkillOrchestra 2026 · VERIFY ORCHESTRATIONOrchRM · MAS-ProVe 2026 · ANALYSISIlluMAS · LiveResearchBench 2026 · ENVIRONMENTEvoHarnessBench TRY IT LIVEMAS-Orchestra Demo NEXT · FRONTIERPolicy · Verification · Env
Our Roadmap

Roadmap of
our work

2025 · Survey
A Survey of Frontiers in LLM Reasoning

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.

TMLR'25🏅 Survey CertificationNeurIPS'25 Tutorial
2025 · MAS Policy
MAS-Zero

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.

inferenceSEA@NeurIPS'25🏅 Oral
2026 · MAS Policy · Analysis
MAS-Orchestra

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.

trainingICML'26
2026 · MAS Policy
SkillOrchestra

Building on the orchestration line — skill-based agent routing, an offline approach that outperforms SOTA RL orchestrators by 22.5% at 700× lower cost.

inference🏅 #2 Hugging Face Daily
2026 · Verify Orchestration
OrchRM & MAS-ProVe

Can we judge an orchestration without running the whole system? Reward modeling and process verification at the orchestration level.

trainingevaluationMAS-ProVe · ICML'26OrchRM · under submission
2026 · Analysis
IlluMAS & LiveResearchBench

When does multi-agent actually beat single-agent? A cost-controlled study plus benchmarks that stress-test the multi-agent advantage.

evaluationICLR'26 · under submission
2026 · Environment
EvoHarnessBench

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.

evaluationtools · skills · agentsunder submission
try it live
Vibe-code your MAS

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.

mas-orchestra.salesforceresearch.ai
“Find all positive integer triples (a,b,c) …”
LEAD SEARCH SOLVE FINAL
Orchestra · designed a multi-agent plan, ready to run
AIMEHotpotQABrowseComp+MASBenchSMFREnterpriseOpsLiveResearchBench
Open the live demo →
What's next
The open challenges

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.

PolicyVerificationEnvironment