Work · IT Solutions
AI Operator
In progressAn internal operator toolkit for orchestrating AI agents across multiple runtimes — one coherent control surface.
- Role
- Internal platform
- Period
- 2025–2026
- Outcome
- Multi-runtime orchestration
03 · AI Operator
Automation that works for people.
An internal operator toolkit for orchestrating AI agents across multiple runtimes — one coherent control surface.
Challenge
The problem.
Agent workloads rarely live in one language or one runtime. Without a shared operator layer, teams rebuild the same controls — start, stop, inspect, route — for every surface.
The need was a single toolkit that remains easy to understand as more runtimes are added, rather than a collection of one-off scripts.
Approach
How we approached it.
01
One operator surface
Runtime differences stay under the hood; operators work through a consistent interface.
02
Polyglot by design
Python core, TypeScript surfaces, Go services and a native client — each where it fits best.
03
Observable by default
Runs, failures and hand-offs are visible without attaching a debugger to every process.
04
Internal-first
Built for our own operators first — clarity over ceremony.
What we built
What we delivered.
Orchestration core
Coordination of agent runs across Python and adjacent runtimes from a shared control path.
Operator interface
TypeScript interfaces for day-to-day inspection, routing and control.
Service layer
Go services for performance-sensitive paths and shared APIs.
Native client
A Swift client for operator workflows that need to live on device.
Internal documentation
Runbooks and presentation material so the toolkit stays usable beyond the original authors.
Architecture
System at a glance.
A polyglot monorepo: Python for the agent core, TypeScript for operator UI and tooling, Go for service endpoints, Swift for the native client.
Shared conventions for configuration and logging keep the stack operable as new runtimes are added.
Numbers & facts
At a glance.
Python · TS · Go · Swift
Languages
Internal operators
Focus
Multi-runtime agents
Scope
In active development
Status
What we learned
What we take forward.
- 01
A shared operator vocabulary matters more than perfect abstraction — teams need the same verbs everywhere.
- 02
Polyglot only pays off when boundaries are explicit; otherwise it becomes four half-finished stacks.
- 03
Internal tools succeed when documentation is part of the product, not a leftover task.
Next steps
What’s next.
- Tighter run inspection and replay for failed agent paths.
- Broader runtime coverage where production workloads demand it.
- Hardened access controls as more teams share the operator surface.
Continue the conversation