Indwel is the enterprise operating system for sovereign cognition. It carries consequential Work from objective to outcome across people, models, and systems—without requiring the institution to reorganize around AI. Models can investigate, reason, and use tools while evidence, policy, decisions, and accountability remain under institutional control.
General availability anticipated in late December 2026
The responsibility for a business undertaking rarely fits within the life of a model session. Whether Work begins as a conversation or arrives from an existing process, Indwel keeps its objective, evidence, authority, history, obligations, actions, and proof together as people and systems contribute. Models, agents, and applications may come and go; the undertaking remains available to those who are accountable for carrying it through.
Adapt AI to the work—not the work to AI.
Businesses already know how to govern consequential work. People operate within roles, use evidence, exercise bounded authority, follow policy, work within budgets, escalate exceptions, and remain accountable for outcomes. Indwel applies the same disciplines to machine inference. Human managers and machine reasoning can participate in the same undertaking without confusing intelligence with authority.
Indwel architecture
Indwel places Work, evidence, reasoning, and consequential action inside one operating system, with institutional authority governing the whole. The same layers appear in the interactive architecture, where the relationships and boundaries can be explored in greater depth.
Indwel OS · canonical architectureExplore the architecture of Indwel OS
ONE UNDERTAKING, OVER TIME
A benefits renewal offers one concrete way to see these distinctions at work, although the operating model is not specific to benefits, procurement, risk, or any other business function. Carrier terms, market benchmarks, formulary details, follow-up questions, committee review, and implementation reach the team on different schedules. Indwel keeps the objective and governing state intact as the record develops, without asking people to treat each new conversation as a fresh start.
Machine inference is one faculty of cognition. Indwel keeps objective, evidence, authority, action, and consequence explicit as reasoning proceeds.
Indwel does not treat fluent generation or agent orchestration as cognition. It can decompose a complex undertaking into bounded cognitive units; deterministically select declared semantic units across native stores and connected data systems; search live and governed sources; compare, calculate, call tools, use specialized agents, verify, repair, and join results under explicit limits and completion conditions. Cognitive units can be tested independently. The structure of the act can remain deterministic even where judgment within it is probabilistic.
What is found does not automatically become fact. Having access to information is not the same as being entitled to treat it as evidence. Indwel preserves the identity and provenance of selected material, tests its standing for the purpose at hand, and admits qualified material as evidence. The same discipline applies to live information from the web and connected systems: source, currency, duplication, corroboration, conflict, and relevance remain visible before a fact may support reasoning.
Finding information is useful, but it is not the same as establishing a factual basis on which the institution can act. Indwel separates retrieved material from admitted Evidence and treats model inference as a contribution to reasoning, not as institutional fact. If the sources are incomplete or contradict the governing objective, the Work can call for clarification or further investigation instead of manufacturing certainty.
The work carries forward what has been established, what remains open, which objective governs, which evidence has standing, and which obligations are still alive. Continuity does not depend on reconstructing an old chat. Long-running cognition can branch into bounded jobs, proceed in parallel, wait on people, time, systems, or external conditions, and rejoin the same undertaking. Agent fleets can carry that work across time while identity, capability, security, budget, evidence, cancellation, and completion remain governed throughout.
A renewal cycle makes the difference tangible: a carrier may clarify a formulary question weeks after the first analysis, a benefits team may refine its objective later still, and the committee may not meet until the evidence is ready. Indwel can keep those developments connected without treating remembered preferences as factual Evidence or giving newly retrieved material standing merely because it arrived later.
Unity is Indwel’s first-party experience. The Platform API exposes the same work, evidence, cognition, memory, authority, action, and settlement model programmatically. The Business App Engine lets domain applications inherit that complete operating substrate rather than assemble their own AI stack and attempt to govern it afterwards.
The simplicity of the experience rests on a substantial operating system. Durable Work and memory coexist with native retrieval, vector-backed persistence, deterministic selection across data systems, traceable Evidence admission, live acquisition, and structured cognition. Governed tools, agents, connectors, and isolated coding environments operate alongside security, inspection, budgets, cancellation, observation, and verifiable Settlement. Enterprise systems can join through governed adapters, with those controls continuing through execution rather than appearing only at the prompt or response boundary.
Models are reasoning resources, not the architecture.
Frontier, specialist, local, private, or air-gapped inference can be selected where the work requires it, used together, or replaced without moving the institution’s evidence, authority, or operating model into a provider. Because no model vendor owns the work, changing providers does not require rebuilding the institution’s operating model or applications. That protects technical and operational investment as the AI market changes.
Governed agents can investigate, use enterprise connectors, write and execute code, and work inside isolated development environments under the same capability, security, budget, evidence, and authority rules as the rest of the undertaking. The applications below are exemplars: different products and interaction models, all inheriting the same Indwel operating system.
REFERENCE APPLICATIONS
Ask the company a consequential question and get an answer whose sources, contradictions, and reliance boundary stay attached. Current policy governs; superseded material remains visible; People Operations—not the model—retains approval authority.
Machine inference can analyze evidence, requirements, contracts, and assurance material without inheriting the right to decide. Here the executed contract contradicts the questionnaire, current assurance is insufficient, and reviewer authority remains explicit.
A healthy build pipeline can establish technical readiness, but it cannot itself grant production authority. The decision to release depends on evidence bound to the exact artifact and an eligible, independent approval; even then, the system records the observed deployment result rather than treating the attempted operation as proof of success.
The same boundary is available programmatically. The application may establish that a release is ready, but authorizing an effect is a separate act bound to the exact artifact, the evidence that supports it, an eligible approver, and a rollback target.
const authority = authorizeReleaseEffect({
disposition: "ready",
releaseId: releaseCandidate.releaseId,
artifactDigest: releaseCandidate.artifactDigest,
evidenceArtifactDigests: [
testReport.artifactDigest,
correctedSecurityScan.artifactDigest,
closureAndApproval.artifactDigest,
],
requesterId: releaseCandidate.requestedBy,
approverId: closureAndApproval.approval.approverId,
approverRole: "production-release-approver",
rollbackTarget: rollbackRunbook.rollbackTarget,
});{ "authorityReceiptId": "auth_4ca72d90", "permittedActions": ["release.deploy", "release.rollback"] }Different experiences can share one institutional operating model without rebuilding evidence, authority, cognition, or proof for every application.
SETTLEMENT
Approval is not effect. An attempted action is not an outcome. Indwel carries consequence back into the undertaking, preserves the authority and evidence that supported it, records what was actually observed, and settles only what the record warrants.
Indwel works across the systems an organization already uses without requiring them to surrender authority or custody. Systems of record remain systems of record. The business process remains authoritative while models, providers, and technical implementations can change around it. Indwel Cloud provides a managed environment for organizations that prefer not to operate the system themselves. Dedicated and licensed arrangements support different custody and operating requirements. Security, inspection, compliance, integrations, and observability attach to the same governed undertaking without turning an AI runtime, provider, or agent into the owner of the enterprise.
Go deeper into architecture, work, platform, integrations, security, deployment, or the compliance atlas.
Bring the Work as it exists today—its systems, evidence, decision rights, actions, and outcomes. Indwel adds machine reasoning without handing truth, authority, or continuity to a model provider. If one consequential process is ready for production design, start with the Governed AI Review.