XSI-AIMS™ v2.0 is public.
Extended Systems Intelligence publishes the XSI-AIMS v2.0 specification for public review — an open governance standard for agentic systems. The comment period opens today. Ratification targets Q3 2026.
XSI-AIMS — the Agent Instrumentation and Management Specification — is a substrate-agnostic, model-agnostic, deployment-agnostic standard for agent governance. It defines four operational altitude layers (L-Intent, L-Plan, L-Form, L-Exec), three functional modes (M-Generative, M-Constraining, M-Integrative), and a conformance principle that separates required capabilities — the spec's WHAT — from delivery mechanisms, the implementer's HOW.
The publication lands at github.com/Extended-Systems-Intelligence/aims: the consolidated specification text, the accepted RFC set, the non-normative companion set, conformance-test material, and the project's governance and IP policies. The text is stable enough to implement against. It is published as a review draft on purpose — the RFC comment period is open, and the spec gets stronger if governance and security teams read it adversarially before v2.0 is ratified.
Two commercial implementations from XSI — XSI-AIMS Atlas™ (a memory routing engine across the XSI-AIMS memory subsystems) and XSI-AIMS Compass™ (a model router with adversarial-challenge for safety-critical calls) — are in progress alongside the public agent registry. They are not finished. That is what the comment period is for.
The research is converging on this architecture.
In the weeks before this publication, three independent research groups arrived at architectural positions that converge with what XSI-AIMS specifies — none citing the spec and, as far as we can tell, none aware of it. A Penn State, UC Santa Cruz, and Amazon team's May 2026 paper (arXiv:2605.30621) defines the agent harness as prompts, skills, memories, and tools — with model weights held fixed — the same plastic-surface thesis XSI-AIMS built its governance model around through this spring's drafting cycle. Two further teams, at CUHK-Shenzhen with Huawei (arXiv:2605.15701) and at ByteDance (arXiv:2605.27366), independently instantiated memory architectures that mirror the XSI-AIMS memory split and promotion patterns, without citing or, as far as we can tell, being aware of the spec.
Convergence is the point. XSI-AIMS does not claim to have invented these ideas — it claims to have synthesized them into a governable contract while the field was still treating them as implementation details. The empirical case is strengthening too: an April 2026 University of Wisconsin–Madison study (arXiv:2604.13151) measured task-success gains of 30 to 37 percentage points from structured harness changes alone, with no weight changes — their result, not an XSI-AIMS benchmark, and direct support for the bet the spec makes.
One claim is deliberately bounded. Among the systems surveyed in our comparative review, none implements prospective memory — future-directed intentions an agent holds and acts on later — as a first-class tier. Published benchmarks measure the in-context failure directly (arXiv:2603.23530); XSI-AIMS specifies it as a core memory type.
"Every major cloud shipped an agent framework this year. They give you tracing and guardrails. XSI-AIMS makes the governance contract itself the product: who authorized an action, under what policy, with evidence you could hand a regulator," said Rhyan J Neble, founder of Extended Systems Intelligence. "XSI-AIMS is our answer, and we're publishing it as an open standard because a governance contract you can't read isn't a contract. We want the hardest critics in the room before this is ratified — not after."
XSI-AIMS is an open standard offered to the larger AI-safety community. Submissions to candidate standards bodies — NIST, IETF, W3C, OASIS, and MCP — are planned in the months following the release. Conformance is all-or-nothing: every required feature present, with how it's delivered left to the implementer. XSI's own products, including XSI LodeStone™ and XSI-AIMS™ for the Microsoft Agent Framework, build against the same published surface.
How to participate: read the spec, file issues for ambiguities, open discussions for design rationale, and submit pull requests against the RFCs. Direct contact for matters that don't fit the public surfaces: via the contact page.
Get the XSI vision essays.
The founder essays on agentic governance and routed-specialization architecture, collected as one whitepaper. Tell us where to send it.