Slide 01 / 09

Welcome to the next era of local agentic compute.

KIRA OS is a prototype agentic desktop environment built around Orchestrator V1. Press Space or click to move closer to the black hole and reveal the system.

KIRA OS + Orchestrator V1Executable local agentic harness, model-in-loop pending
75 / 100
Codex CLI + GPT-5.4Harness-Bench published agent-harness result
80.4
Hermes AgentHarness-Bench aggregate harness result
71.2

Professional reference: Harness-Bench evaluates 106 sandboxed workflow tasks with full agent execution traces. Published results report Codex at 80.4 using GPT-5.4 and Hermes Agent at 71.2 as an aggregate harness result. KIRA's 75/100 is from an executable local prototype harness that ran KIRA pathways for PPTX, DOCX, PDF, permission queueing, virus scan, optimisation scan, artifact indexing, and reasoning stripping. Orchestrator model-in-loop and official Harness-Bench are still pending.

Space or click moves through the orbit.

Built as a local agentic environment, not a chatbot skin.

Orchestrator chooses the route. KIRA OS provides the computer pathways, permission checks, generated artifacts, live context, and proof layer.

01 / context

Computer awareness

Read-only scans for apps, files, windows, RAM, storage, downloads, project state, and live computer context.

02 / create

Artifact generation

Fresh PPTX decks with images, DOCX reports, PDFs, notes, package manifests, and research summaries.

03 / act

Desktop pathways

Open apps, browse the web, inspect pages, connect IDEs, draft MCP configs, and prepare automation steps.

04 / guard

Safety boundary

Read-first behavior, permission-gated mutation, macOS sandbox direction, protected-path blocking, and failure reporting.

05 / verify

Evidence layer

The agent should return file paths, browser state, command results, screenshots, or logs before claiming completion.

06 / scale

Council and background work

Confidence-based council mode, background tasks, live mode, and multi-step recovery for hard agentic workflows.

Agentic benchmark signals.

KIRA OS now includes an executable local agentic harness score, while Hermes and Codex use published professional agent-harness reference results.

KIRA OS + Orchestrator V1Executable local agentic harness, model-in-loop pending
75 / 100
Codex CLI + GPT-5.4Harness-Bench professional result
80.4
Hermes AgentHarness-Bench aggregate harness result
71.2

Harness-Bench is a professional benchmark of 106 sandboxed workflow tasks across software engineering, data analysis, workspace/tool operations, knowledge retrieval, office communication, vertical workflows, long-running autonomy, and DevOps. The paper reports Codex with GPT-5.4 at 80.4 and Hermes Agent at 71.2 as an aggregate harness result. KIRA OS scores 75/100 on its included executable local harness, which actually runs local KIRA pathways where possible and reserves points for Orchestrator model-in-loop plus official external Harness-Bench work that was not run here.

Prototype release.

KIRA OS is a local agentic environment for Orchestrator V1, packaged around desktop pathways, generated artifacts, evidence, and safety-aware execution.

Prototype version. Orchestrator V1 is public on Hugging Face, and the KIRA OS prototype code is available on GitHub.

Follow the author.

Follow Saggamer for KIRA OS updates, Orchestrator V1 releases, local agentic compute experiments, and prototype demos.