Platform · Orchestration
Orchestration agents
A modern inertial sensor is not one problem but five coupled ones. Orchestration agents put a supervisor over a team of specialists so a single design change is reconciled across every physics before anyone trusts a number.
01The problem
A multi-axis inertial-MEMS redesign of the kind Bosch Sensortec and STMicroelectronics ship couples electrostatic actuation, structural mechanics, squeeze-film damping, thermal stability and package stress in one part. Move the drive combs to raise gain and the change ripples straight into sense-mode frequency and quadrature error, so no domain can be tuned in isolation. Run it as five separate studies and the interactions that actually set yield stay unmodelled.
02How we solve it
A supervisor over specialist agents.
Agentriq decomposes the redesign into a task graph and routes each sub-task to a specialist agent that reads and writes one evolving state.
Supervisor and hierarchical routing
A supervisor agent holds the global state, decomposes the redesign and delegates each sub-task to the worker best suited to it, managing every hand-off between them.
Planner-executor (plan-and-solve)
A planner turns the goal into an ordered plan up front and an executor runs it step by step, cutting the redundant reasoning of improvising one action at a time.
Hierarchical task decomposition (task DAG)
The redesign is expressed as a directed graph of sub-tasks, so independent branches for thermal, damping and electrostatics run in parallel instead of one slow chain.
Adaptive, state-triggered replanning
When an executor result shifts the shared state, such as a new sense-mode frequency, the planner repairs only the affected branch rather than re-running the whole plan.
03What it produces
What per-domain studies miss.
One reconciled design pass where each drive-comb change is checked against sense-mode frequency, quadrature and damping at once, surfacing the coupling a per-domain sum hides: in one published gyroscope study, fully coupled analysis put drive-mode Q near half the value the summed-mechanism estimate predicted.
04The agentic loop
Inside the closed agent loop.
Agentriq wraps the supervisor-and-specialists pattern in a closed loop that plans, dispatches, verifies against the coupled model, replans on conflict, and lets a judge decide before the state is written back.
- 01The supervisor decomposes the redesign into a task DAG and routes sub-tasks for electrostatics, structure, damping, thermal and package to specialist agents over a shared state.
- 02Specialists run their tools, COMSOL and Ansys, so a change to the drive combs propagates immediately into sense-mode frequency, quadrature, damping and package stress.
- 03Every candidate is verified against physics checks and the fully coupled model, flagging where a per-domain estimate disagrees with the coupled solution.
- 04On a conflict, the agent traces root cause with Agentic RAG over prior runs, datasheets and literature, then replans only the affected branch.
- 05A judge scores the reconciled design against targets and constraints, then writes the accepted state and its rationale back to memory for the next loop.
05Tooling
Simulation, solvers & frameworks.
Get started