We build our own agents before we build yours.

Agent runtime

A self-hosted runtime for agents that are allowed to touch real systems.

Harbor runs agent workloads inside a sandbox with explicit tool permissions, full request tracing and an immutable audit log. It is what we deploy when a client needs an agent inside their own network and their security team needs to read every action it took.

• Per-tool permission grants with human approval steps
• Immutable audit log for every tool call and model response
• Deploys into a client VPC with no outbound data path by default

Evaluation harness

Model and agent evaluation that runs in CI, not in a meeting.

Sift turns quality into a versioned number. It holds datasets, adversarial suites and regression thresholds, runs them reproducibly against any model or agent version, and blocks a release that scores below the line.

• Versioned datasets with reproducible, seeded runs
• Adversarial and long-context suites built per capability
• Regression gates wired into existing deploy pipelines

Voice agent

A voice agent for operations teams that hand off cleanly to a human.

Relay handles scheduling, intake and escalation over the phone in Arabic and English. It scores its own confidence on every turn and routes anything below threshold into a human review queue with the full transcript attached.

• Arabic and English with per-tenant vocabulary
• Confidence scoring with automatic human handoff
• Writes back to the systems of record, not to a scratch database

Embodied agents

Perception and task planning for robots working in unstructured sites.

Tessera is our internal research track on embodied agents: grounding language instructions in a live scene, planning multi-step physical tasks, and recovering when the world does not match the plan. Not a product yet. We publish what we can.

• Scene grounding from monocular and depth input
• Multi-step task planning with recovery behaviours
• Simulation-to-site transfer testing