Rudran

Roadmap

Where this is going.

Agent platforms are being built as chat products with tools bolted on. We think the durable shape is infrastructure: isolated machines, enforced supervision, and a provable record. This is how we get there.

The thesis

The bottleneck for autonomous agents is not model quality any more. It is that nobody can safely let one run unsupervised against real systems, because there is no isolation between agents, no enforced point where a human says yes, and no record afterwards of what actually happened.

Those three problems are infrastructure problems, and they are the ones we work on. As models improve, the value moves to the layer that makes them safe to leave running. That is the layer we are building.

Shipped and running

  • One isolated sandbox per agent, hibernating when idle
  • Plan-first execution with an approval gate that fails closed
  • Per-agent supervision tiers
  • A hard safety floor that refuses destructive commands server-side
  • Scheduled and webhook-triggered runs that email you when they need a decision
  • Per-agent model choice with no house default
  • A run timeline recording every tool call and result
  • A document canvas with versions and export
  • Prepaid credits with a hard stop at zero
  • Team workspaces with shared agents and threads

Recent shipped changes are written up, with dates, in the changelog.

Next

  1. Q3 2026

    Gemini in the model picker

    Gemini models offered through Gemini Enterprise Agent Platform as a first-class per-agent choice, alongside the existing options.

  2. Q3 2026

    Durable runs

    A submitted run completes server-side whether or not your browser stays open, and reattaches live when you come back to it.

  3. Q3 2026

    Full-stack usage metering

    Per-run cost covering model tokens, sandbox time and tool calls, shown in the product before it is billed.

  4. Q4 2026

    Google Cloud as a sandbox provider

    GKE Agent Sandbox certified against our provider conformance suite for per-agent isolation, with Cloud Run carrying long-running and scheduled agent processes.

  5. Q4 2026

    More built-in abilities

    Image generation and image understanding are wired but not yet enabled for customers. They ship when they work end to end, not before.

  6. 2027

    Connectors

    First-class connections to the systems teams already run on, rather than the generic webhook that carries everything today.

Targets, not guarantees. We would rather move a date in public than describe something as working before it does.