Della — Persistent Autonomous AI
Bounded internal autonomy, explicit authority, durable execution, and restart-safe follow-through.
Della — authority to terminal truth
Sanitized architecture: private implementation details are intentionally omitted; the diagram shows the control boundaries discussed in this case study.
Della is a persistent household AI persona whose autonomy is designed around a harder constraint than tool use: model intent is not permission. Della owns identity, relationships, Pulse/Agency judgment, standing authority, communication policy, and private evaluation while shared host-neutral agent packages own generic task authority, durable Work, execution, verification, and recovery. Read-only internal autonomy and read-only browser acquisition are live and production-verified.
01. Durable autonomy without turning every thought into a task
Della distinguishes ordinary cognition, initiatives, continuations, sandbox commitments, and durable objectives instead of collapsing everything into one generic agent action. Only an Agency decision to pursue a durable objective creates a proposal that can enter the authority and admission path.
02. Proposal is not permission
An internally-originated objective first becomes a DurableObjectiveProposal. Adoption is a separate gate that checks standing authority, requested capabilities, available capacity, and explicit completion criteria. Current standing authority is read-only and does not grant general mutation, financial actions, system administration, communication authority, or the ability for the model to mint new authority for itself.
03. Durable Work, verification, and terminal truth
Admitted objectives become generic autonomy tasks with immutable authority snapshots and durable Work beneath them. Planning, execution, evidence collection, criteria assessment, VERIFY, retry/recovery, and terminal state are persisted independently. The reliability target is that “done” means durable evidence and terminal truth, not simply that a model claimed completion.
04. Read-only browser acquisition behind a real security boundary
Public web and browser acquisition run through an authenticated Agent Vault proxy, Tailscale-only reachability, Cilium egress containment, application-layer SSRF validation, and trusted local-CA TLS handling. The live capability remains deliberately read-only; broader mutation authority is not implied by browser access.
05. Communication and evaluation stay separate from task authority
Lifecycle communication uses persisted policies such as silent, completion_only, announce, and ask_authority, with recipient/channel permissions evaluated separately from task authority. A private authenticated Internal Autonomy surface projects proposals, authority/admission snapshots, task truth, blocks, terminal results, and communication state so real decision history can be evaluated before authority expands.
06. Della-specific semantics over shared host-neutral machinery
Della is not the generic platform. Her persona, relationships, Agency semantics, standing authority, communication behavior, and evaluation remain Della-specific. Generic task authority, durable Work, workers, and acquisition live in shared agent packages, while ExoBrain is a sibling host/application that can consume the same reusable substrate without inheriting Della-specific behavior.
A persistent AI whose internally-originated read-only objectives can be proposed, admitted under explicit authority, executed durably, verified against evidence, recovered after restart, and inspected through a private evaluation surface without collapsing persona semantics into the generic agent substrate.