Per-turn model routing across quality, cost, latency, capability, cache state, and reliability objectives
Provider and protocol portability across frontier models, open-source models, local inference, and compatible APIs
Agent and harness interoperability, including transferable skills, capability profiles, actions, tools, and trajectories
Portable, user-owned memory and context with scoped identity, provenance, retrieval, feedback, and reviewable compaction
Agent interchange standards, conformance testing, tool and MCP access, and agent-to-agent communication
Agent and harness optimization through evaluation, distillation, customization, and multi-agent learning
You will design and build research prototypes and robust systems at the seams between models, providers, and agent runtimes. You will formulate research questions, develop evaluation methods, test ideas in realistic agent workflows, and turn promising results into reusable components. The work will often involve collaboration with adjacent research, infrastructure, security, product, and engineering teams, where findings are validated and applied in practice
We are currently looking for senior- and staff-level ML engineers to work on research in areas such as
Learned, rule-based, and hybrid model routing, cascading, and candidate-ranking systems
Quality-cost-latency trade-offs, uncertainty estimation, exploration, and outcome-aware routing
Multi-provider gateways, protocol translation, catalog normalization, and fail-closed execution contracts
Portable agent skills, harness capability discovery, package adaptation, and cross-harness conformance
Memory, identity, context, trajectory, and outcome representations that remain portable across agents and models
Retrieval, context selection, context compaction, and feedback systems with explicit provenance and trust boundaries
Agent interoperability standards, including metadata, action formats, plugins, tools, MCP, and agent-to-agent interfaces
Agent optimization, teacher-student distillation, skill generation, harness customization, and multi-agent learning
Benchmarking and evaluation infrastructure for model, router, memory, skill, and harness changes
Designing, implementing, training, and evaluating model routers that select an appropriate model or reasoning profile for each turn
Developing portable provider and protocol abstractions that preserve authentication, telemetry, cache and context signals, and execution provenance
Defining versioned schemas and contracts for models, provider offers, agents, workspaces, skills, actions, tools, memories, and trajectories
Building systems that discover, package, adapt, and validate agent skills across coding agents, editors, and other harnesses
Researching user-owned memory, scoped identity, trajectory checkpoints, terminal outcomes, retrieval quality, and reviewable context compaction
Creating benchmark suites and evaluation protocols for quality, cost, latency, reliability, safety, and portability
Designing held-out, out-of-domain, and change-impact evaluations that test new or removed models, providers, skills, and harness versions
Investigating distillation, self-improving harnesses, multi-agent training, agent factories, and automated skill creation
Writing robust research software, APIs, integration layers, and test infrastructure that enable rapid but reproducible experimentation
Collaborating across research and engineering teams to translate promising ideas into secure, reversible, and reliable systems
Communicating results through technical reports, demonstrations, open-source releases, benchmarks, and research publications