Changelog

What shipped.

Changes to arkenOS, Optimus, and the loop — dated, specific, and in order. No roadmap promises.

AUG 3, 2026

The research workspace convention

A dedicated research root. Every project now carries a conventional research directory that the workshop, the agent, and the API all agree on — one place where specs, runs, and evidence live.
Stable identities for every type. The formally specified types that govern a workspace now resolve by permanent identifiers, so a spec written today validates identically years from now.
Research leadership. Optimus was recast as a research leader with progressive skill routing — each specialist scoped to exactly one research concern.
Auditable search. Web and private-knowledge search are available to the agent with sealed, auditable activity — every source a conclusion used is on the record.

JUL 31, 2026

A durable graph engine for training campaigns. Runs execute as graphs with per-node state, topology and contract validation, and a full canvas UI: library, workbench, inspectors, and live execution views.
A catalog of ready-made campaigns. More than two dozen bundled patterns — train-and-gate, promotion gate, experiment matrix, scenario builder, dataset curation, replay reproduction, environment factory.
The specialist crew. Sixteen named specialists, each with a defined role, shared working principles, and per-specialist persistent memory — including the adversarial roles that attack the others' conclusions.
Goals, plans, and memory. Durable goals authored before activation, session plans, turn recovery, and cited recall — the agent can tell you where a belief came from.
Teach sessions. Human-in-the-loop capture: when the crew gets stuck, you drive, and the demonstration becomes training signal.
Checkpoints and rewind. Conversations fork, rewind, and resume; edits backtrack the code with them.
A managed desktop runtime. The full platform — services, storage, runtime — supervised inside one installable application.

JUN 30, 2026

The execution plane moved fully local. Cloud pods retired; compute, telemetry, and the work queue rebased onto machines you own.
Optimus took residence. The agent got its own runtime, its own persona, and a place in every workspace turn.
A real workspace file layer. Files shared natively between the platform and the agent, with external-editor support and desktop integration.
Chat that keeps up. Token-level streaming on a direct local path, and durable conversation history that outlives the process.
Pages and viewers. Rich workspace pages with MDX rendering, data and media blocks, document viewers, and page templates.

MAY 31, 2026

Experiment tracking, end to end. Every engine run logs its metrics, parameters, and artifacts automatically; a run browser puts the full history behind one surface.
A GPU job fabric. Dependency-aware scheduling with retries, budget accounting, and recovery — training jobs queue, execute, and leave logs and artifacts behind.
Evaluation gates. Promotion became evidence-backed: a policy moves forward only with its gate report attached.
Patch review. Changes an agent makes to a workspace arrive as reviewable diffs with test results — nothing lands unseen.
A kinematics lab. Forward and inverse kinematics, manipulability diagnostics, and trajectory authoring for embodied models, with reachability and joint-limit warnings.

Work with us

We are taking a small number of design partners. The bar is a real workload, not a logo. Everything runs on your hardware, and you keep the machine.