Model Behaviour Architect - General Agents & Tooling

Mistral AI; Greater London; On-site; GBP 110,000 - 170,000; Full time

Benefits offered by this job

Healthcare

Parental leave

Retirement plans

Relocation support

Wellness programs

Meal and transportation allowances

Job summary

Mistral AI is seeking a Model Behaviour Architect to define and measure how large language models interact with tools, invoke functions, and orchestrate workflows on the General Agents team. You will ensure models perform accurate, reliable tool use and robust multi-step orchestration.

Collaborating with our Science team, you will turn research insights into concrete evaluation pipelines and guidelines, shaping what 'good' looks like for function calling and enabling greater autonomy and

Qualifications

• Strong knowledge of API design and JSON Schema.
• Experience with LLM agents, reasoning, planning, and multi-step tool use.
• Experience training or optimizing model behaviour for real-world apps.
• Proven ability to build scalable evaluation frameworks.
• Ability to work in dynamic, technically complex environments.

Responsibilities

• Identify opportunities to improve function calling and tool use in models.
• Gather feedback to scope enhancements for model behaviour.
• Design evaluations, data guidelines, and synthetic tool environments.
• Address edge cases and perform rigorous testing.
• Develop evaluation pipelines for model candidates' function-calling.
• Collaborate with AI Scientists to align behaviour with research insights.

Skills

API design & JSON Schema

LLM agents & tool use

Reasoning & planning

Model behaviour tuning

Evaluation frameworks

Job description

Mistral AI is seeking a Model Behaviour Architect to define and measure how large language models interact with tools, invoke functions, and orchestrate workflows on the General Agents team. You will ensure models perform accurate, reliable tool use and robust multi-step orchestration.

Collaborating with our Science team, you will turn research insights into concrete evaluation pipelines and guidelines, shaping what 'good' looks like for function calling and enabling greater autonomy and