Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop?

What it claims

Mecatl is an open-source agent harness that Stacklok started on GitHub in June. It is built to run in the cloud from the start, instead of a desktop harness moved into a VM or container. Beda's argument is that "lift and shift" breaks agent lifecycle management, especially quiescing: safely pausing an agent while it waits for human input. Desktop harnesses, even pluggable ones, run the loop, local execution and session state as one process or tightly coupled processes. That is what prevents clean pausing and central management.

How it works

Mecatl keeps the agent loop independent of four things: the client, the model provider, the state store and the execution environment. The loop is treated as an ordinary cloud application. Sensitive operations such as tool calls and bash run separately. Session state and memory are stored in "manageable systems" instead of JSONL files on disk. The intent is for enterprises to own and govern agent interactions the way they govern email.

flowchart LR
  C[Client] --> L[Agent loop]
  L --> M[Model provider]
  L --> X[Isolated tool / bash execution]
  L --> S[Session & memory store]
  P[Enterprise spine: identity, policy, audit] -.governs.-> L
  P -.-> X

ToolHive began as a way to manage MCP servers running in Docker containers on the desktop. It then grew a Kubernetes-based gateway, a registry and an operator, adding I/O and auth handling for MCP servers.

How to use it

The page gives no install commands, licence names or API details. The useful information is where the open-core line falls:

• Open source: Mecatl and ToolHive. Each is meant to be useful on its own and "relatively easy" to run for a small team.
• Not yet open source: the AI Gateway, which provides access control, budgets, reporting and provider routing. Beda says it is on the roadmap.
• Commercial: the "enterprise spine," a control plane that provides consistent identity, authorization, policy and auditing across the open-source pieces. It targets multi-cluster and multi-cloud deployments.

Stacklok's customers are mostly banks, semiconductor companies and telcos. The company raised a $17.5M Series A in 2023 from Accel, Madrona and Bain Capital.

Limits

• No task-based model routing. The AI Gateway does not choose a model per task; the page contrasts this with Glean's product. Stacklok's position is that semantic routing belongs in the harness, which has more context. If you need routing by task, the gateway will not do it today.
• Scaling features are paid. Multi-cluster and multi-cloud operation is where Stacklok places its commercial product. The open-source projects alone may not cover large-scale governance.
• Vendor claims only. The page has no benchmarks, reliability data or comparison against harnesses like Pi. Beda did not address Pi directly. The case for Mecatl rests on architectural argument and the Kubernetes analogy.