← Back to Experience

Connectra: how an agent step runs

Cofounder & CTO, Sep 2025 – Present. The agentic platform I'm building, drawn as a machine you can run.

The job card, whatever is being checkedThe trace each station drawsWhere the seal landsOne agent step, drawn as a machine. A card goes to the drawer for its tool, is checked at the gate against its schema and sent back if a field doesn't fit, and is sealed by a press that only turns with the agent's own key. The recorder in front draws what every station did. An illustration of the design, not measured data.

Orchestration

Multi-step LLM orchestration pipelines with tool-calling and structured JSON outputs, so that every agent decision is consistent, machine-readable and auditable end to end. In the figure this is the tool rack and the gate: the card goes to the drawer for its tool, and only output that fits the schema is allowed to move on.

Observability

An observability layer that traces agent decisions, actions and outputs across distributed workflows. In the figure this is the recorder: every station has a pen, and every event leaves a mark on the paper.

Identity

Agent identity and credentialing, so that every action can be attributed to a specific agent. In the figure this is the press: it only turns with the agent's own key, and the seal it leaves carries the same pattern as the key.