Agent & Pod
The Agent is a replaceable execution character. The Pod is a disposable machine-side realization. They are different layers of recovery.
Developer Preview ↗AUTONETES / PLATFORM
Developer Preview / qualification in progress ↗One operating model.
From intent to evidence.
The WorkUnit holds the mission. Autonetes governs who can act, what survives replacement, and why the result can be accepted.
Separate the work from the Agent, the machine-side Pod, and the intelligence and runtime serving it. Recovery preserves the mission’s authoritative record.
The home for the objective, scope, current authority, checkpoints, effects, delegated work, evidence and completion.
The Agent is a replaceable execution character. The Pod is a disposable machine-side realization. They are different layers of recovery.
A versioned specialist suit of skills, knowledge, interfaces and requirements. Specialist capability never grants authority by itself.
A legal space to navigate rather than a rigid sequence. Design defines CAN GO; current projection defines MAY GO; traversal records DID GO.
RoamLooper addresses progress and semantic stalls; its efficacy requires empirical qualification. PathProof records lawful traversal, not the correctness of every inference.
A new execution.
The same authoritative WorkUnit.
A restart alone is not recovery. A successor needs valid continuity, fresh authority and reconciled effects. Recovery should occur at the lowest safe layer.
Inspect the authority boundary ↗A coherent operating environment over the same public Core semantics used by the CLI, SDK and API. These views illustrate the product model.
Set the objective, scope, specialist capabilities, legal navigation and completion requirements before execution begins.
CORE SEMANTICSAn Agent requests a specialist. Core admits a Child WorkUnit with its own scope, fence, Shute and qualified execution. Child authority cannot exceed ancestral delegable authority.
Agents execute. Work delegates. Autonetes initiates.
Admitted child work can survive parent Agent replacement. A legal delegation node permits a request; it never automatically spawns work.
A model can propose. Consequential actions require current authority. Uncertainty stays visible until evidence establishes what happened.
Toggle the required conditions for a scoped report write.
A missing acknowledgement does not establish failure.
An acknowledgement is recorded. Explore what changes when it is lost.
This illustration uses a confirming downstream probe. Real certainty depends on downstream semantics and available evidence. No universal exactly-once guarantee.
Verification accepts a declared outcome against evidence. PathProof establishes lawful traversal. Actual business impact is measured independently.
These examples describe potential applications, not customer deployments or qualified integration claims. Each mission has explicit boundaries.
Coordinate ingestion, market validation, anomaly analysis and independent verification to produce an evidence-backed investigation package.
Explore 13 industries and domains—including healthcare, logistics, manufacturing, marketing, insurance and the proposed Physical AI extension—with explicit authority boundaries and evidence requirements.
Explore industries & domains ↗Operate through Mission Control, automate through the CLI, build through the SDK, and integrate through the API. Every surface uses the same public Core semantics.
Developer Preview status ↗autonetes work delegate …ctx.spawn_subagent(…)DelegationRequestConceptual surface examples, not installation instructions. Supported APIs and distribution will be published against qualified preview artifacts.
Models provide qualified intelligence. Runtimes provide qualified execution. Business systems remain systems of record. Autonetes does not replace Kubernetes, IAM, secret management or observability.
Autonetes is being qualified for bounded autonomous workloads. Project materials report substantial Core foundations. Distribution, supported adapters and exact release claims must be qualified against specific artifacts.
Explore the company direction ↗Preview themes include worker replacement, governed models and effects, delegated work where qualified, and proof inspection.
The operating model, its boundaries, and the current product stage.
Autonetes is building a unified operating foundation for autonomous work: Agent lifecycle, durable work, current authority, governed effects, execution supply, recovery, accountability and cost visibility. The enterprise defines its objectives and decision rights; Autonetes is designed to make that work-level contract operational across qualified components. Maturity varies by capability.
The latest company materials report internally qualified Core foundations around durable WorkUnits, current fences, effect uncertainty, checkpoints, succession and evidence. An independently usable installation and SDK, Mission/Release readiness experience, qualified external effect profiles and operator integrations remain productization milestones. A shipped Developer Preview, Enterprise GA, production SLA, signed design partners and customer revenue are not claimed.
In the target operating model, a standing Mission defines a repeatable business responsibility. A Mission Release defines and qualifies its operating contract. Individual WorkUnits are bounded cases admitted under that contract, each with its own current authority, effect history and completion record. One Mission can serve many WorkUnits; it is not one immortal Agent or one root WorkUnit. Standing readiness remains a productization target, not a guarantee that every case succeeds.
A policy decision is insufficient if execution can bypass it. The design requires credentials, egress, gateways and qualified adapters to enforce the same current Scope, Calibration, Fence and approval obligations at the effect boundary. Strict profiles must fail closed when currentness cannot be established. Enforcement depends on the actual qualified path; it is not a claim of universal coverage across external systems.
The WorkUnit retains committed progress, eligible checkpoints, authority history, the effect ledger and evidence. Succession advances the fence and supplies fresh current authority; stale execution must be refused at the enforced effect boundary. A missing acknowledgement leaves an effect UNKNOWN, not FAILED. Recovery reconciles available evidence before deciding whether a retry is safe; arbitrary external effects have no blanket exactly-once guarantee.
Completion is a governed decision against declared acceptance requirements and recorded evidence, rather than an Agent’s own status report. It accounts for required child outcomes, effect certainty and applicable verification. PathProof records lawful traversal, not correctness of every inference. Accepted work is also distinct from business impact: incremental value and all-in cost per independently accepted outcome must be measured against an appropriate business baseline.
No. Models, Agent frameworks, runtimes, clouds and business applications retain their roles. Autonetes is designed to provide a common work-level contract across qualified components. Business systems remain the source of committed business truth. Provider freedom is a design goal; specific adapters, capabilities and execution paths require qualification.
Begin with one consequential, bounded operation: a business owner, approved source records, explicit permissions and human gates, qualified effects, and acceptance criteria. The latest commercial hypothesis focuses on freight recovery, from an incoming charge through investigation, approval, governed dispute and reconciliation. That is a proposed starting wedge—not an installed vertical package or a signed customer deployment. The test is whether a second distinct operation can reuse the same Core contract.