Released: an agent that stays with you, with disposable work cells

[AlexsJones]

An agent that stays with you. Small work cells that come and go.

We’ve released Sympozium v0.10.57 and Celln v0.5.8, bringing persistent native agents and disposable work cells together.

The idea is simple: keep the conversation alive, but give each piece of work a fresh, tightly limited place to run.

What does that mean?

Imagine an agent helping you work through a task over several messages. You want it to remember what you just discussed. You also want to choose what it can read, write or fetch—not hand it the keys to everything.

A parent cell keeps the native agent’s live context. For each turn, it uses a disposable child cell—a “sub-cell”—with bounded authority. The work finishes, its result comes back, and the child is cleaned up. The parent stays available for your next message.

There’s something almost biological about it: continuity in the parent, short-lived specialist work in the children. It’s an analogy, not a claim that the agent reproduces or grants itself new powers.

For a one-off job, you can still use a disposable cell on its own. That job doesn’t even need an AI model: it could be a deterministic tool invocation.

Why we’re excited

An agent’s lifetime and its access to tools no longer need to be the same thing. A continuing conversation needn’t mean giving every turn a permanently open toolbox.

Sympozium remains the place where you choose the agent, manage policy and see runs. Celln provides the hardware-isolated execution environment. The first native toolbox supports workspace read/write and allowlisted HTTPS fetch, with model credentials held on the host. Suggested defaults are choices, not automatic permission grants.

In plain English: the harness is how the agent runs; skills provide instructions; tools are the actions it may take; Celln is where this native work executes. Existing Kubernetes agents and OCI harnesses remain available too.

This is running, not just a diagram

On real KVM hardware, we tested a real DeepSeek-backed conversation retaining context across turns, borrowed tools, controller upgrades and normal cleanup. We also ran the existing Pi and Hermes persistent-harness regression suites successfully. The final released native controller recalled the earlier word correctly, then cleaned up the run normally.

This first release is intentionally focused: Linux amd64/KVM, a single owner, one active turn and a small live context budget (about 2 KiB). “Persistent” means the parent stays alive between turns—not that its context or files survive a host crash. Arbitrary OCI harnesses, Python and shell inside native cells, checkpoints and parallel turn trees are not included yet. The release notes document these limits and a separate failing clean-Kind recall test under investigation.

Try it and help shape what comes next

Start with the native installation guide. It covers the operator setup and approved assets needed before selecting the native Harness, Celln and enduring lifecycle through the UI or YAML. This is an opt-in path, not a change to everyone’s default agent.

The delivery epic tracks the work beyond this first release.

What would you lend an agent first? Which task would become more useful if the conversation stayed alive while each turn’s work stayed tightly contained? Bring examples, questions and objections here—this is the new home for the combined feature discussion.

How we got here

These earlier conversations shaped the release and remain useful background:

• External execution plane: Kubernetes as coordination plane—separating the place that manages work from the place that executes it.
• Persistent agent APIs on Celln—exploring a resident cell versus short-lived cells and explicit state.
• AgentHarness: current state, security boundary and next steps—bringing approved harnesses into Sympozium without losing platform controls.

This release brings those threads together without replacing their broader architectural questions.