UNDERSTANDFriday has one hard constraint.
1/4WHISPER / INITIATIVE UNDER DEVELOPMENT
TRULY
HUMANAUTOMATION.
Whisper is being developed to turn everyday conversations into completed, verifiable outcomes—through permitted applications, devices and services, with authority you control.
Friday has one hard constraint.
Your 4:30 call cannot move. This example connects a message, calendar, location and flight options.
Your 4:30 call cannot move. This example connects a message, calendar, location and flight options.
An agentic operating layer for everyday life.
The proposed operating layer combines personal context, bounded planning, device and service execution, and independent verification. The goal is less life administration, with the important decisions and permissions still in your hands.
No app required
The intended experience uses text, calls and voice notes through approved channels, with an optional dashboard for inspection and control.
Uses devices like you do
The design combines supported APIs with isolated computer or phone operators, using explicit permissions for screen, audio and local software.
Finishes life admin
The proposed specialists coordinate appointments, purchases, negotiations, forms and service providers against observable completion conditions.
Its own email
A planned dedicated address separates authorised registrations and merchant threads from the personal inbox, with identity and send controls.
Maintains an active state
Permission-aware temporal memory is designed to retain relevant commitments, preserve contradictory evidence and identify material changes.
Acts inside your authority
The authority design binds scope, expiry, spend and consequence to each action; independent external-state read-back is required for completion.
Conversation-first designText and voice through approved channels.
Adaptive agent executionSpecialists re-plan across devices, services and changing conditions.
Independent completion proofCompletion requires the external state to match the contract.
Text it. Call it. Keep moving.
The experience is designed around one permission-aware task state across text, voice and approved messaging channels, with context and control carried between them.
Whisper meets you where conversation already happens.
The intended experience supports text, calls and voice notes through approved channels. Reactions acknowledge context; consequential actions need explicit, action-bound approval.
It uses software the way a person would.
The integration design covers email, calendars, messages and explicitly permitted screen, audio and location context. It prefers supported APIs, with isolated interface operators for authorised legacy workflows and human handoff at authentication or platform boundaries.
Find and book an in-network doctor
Thursday afternoon · within 8 km · existing insurer
Context resolved across permitted sources
Whisper separated hard constraints from preferences and located every credential, record and conversation needed for this outcome.
What does success mean?
Resolve the real outcome before touching a tool.
- Authority
- Scoped execution
- Exposure
- Credentials isolated
- Memory
- Outcome + proof
Give the work an inbox of its own.
The planned agent inbox separates operational communication and supports scoped thread handover. A CC provides context, not authority to accept terms or send sensitive information; mailbox identity, retention and approval rules remain explicit.
An agent-owned address. A quieter personal inbox.
The planned address supports permitted account administration and merchant communication. Thread access and external-send authority are separate permissions.
Real work, carried through to reality.
The intended workflows connect research, conversation and software to external service providers. Submission, provider acceptance and actual completion are different states; unknown or contradictory results must remain open for reconciliation.
Life fits together, not into separate apps.
Doctors, DMV appointments, restaurants, flights and road trips are planned against calendars, coverage, location and the people affected.
Coverage checked
Whisper recovered the people, timing, preferences, policies and previous conversations relevant to this outcome.
It remembers the promise, not just the prompt.
The memory design retrieves relevant, fresh and permitted evidence rather than treating every past message as current truth. Proactivity is opt-in, with quiet hours, notification limits, inspectable inferences and separate approval for consequential actions.
The right intervention, before the open loop becomes a problem.
Whisper remembers commitments, watches permitted signals and surfaces only what has become consequential.
WHY THIS MATTERS NOWYou promised Sarah the risk note.
DetectedCould you send the final risk note before the board brief?
READY ATTACHMENTBoard-risk-note-v7.pdf
FINALWhisper linked the promise, final file and tomorrow’s dependency.
The file is complete, but the conversation has been quiet for four days and tomorrow’s board brief depends on it.
The same intelligence layer. Mission scale.
WhisperPrime proposes a separate deployment profile for authorised intelligence analysis, government coordination, security and critical infrastructure. It combines source-governed queries, evidence-linked assessments and named human release within deployment-specific sovereignty and assurance controls.
Maritime OSINT
STIX/TAXII + typed claims
- Live state
- 18 corroborated
- Admitted objects
- 31 entities
- Handling
- OPEN · REL PARTNER
- Custody
- Owner-controlled
- Interface
- Delta / S3 query-at-source
- Query mode
- Read-only claims
Purpose-bound query enforcedNo source grants broader access than this mission purpose requires.
Sovereignty includes inference
Query-at-source alone is insufficient. Specify data residency, model and embedding location, key custody, administrator access, support paths and log placement. Derived intelligence retains the source's handling and releasability constraints.
Federation requires semantic agreement
Version entity and event schemas; reconcile identifiers, units, timestamps and lineage. Enforce purpose, role, mission and time attributes at each source gateway; handle policy conflicts and disconnected operation explicitly.
Assurance is deployment-specific
ABAC, mTLS, STIX/TAXII, OGC and OPC UA are design choices, not proof of accreditation. Keep OT access read-only until independently approved, disclose inference residuals and require named human release of consequential mission actions.
Federated interoperabilityModern and legacy platforms connect without collapsing data boundaries.
Mission-governed autonomyPurpose, role, policy and time bound every agent action.
Evidence-linked intelligenceClaims retain source, confidence, transformation and handling rules.
Cross-agency executionStakeholders coordinate through shared outcomes and accountable gates.
More capable. Still yours.
Task-scoped access
The proposed gateway leases credentials and tools only to a bounded outcome.
Consequence-aware approval
Purchases, binding commitments and sensitive changes remain gated.
Control at any moment
Pause authority, revoke connections, delete memory or stop execution.
Proof, not promises
Completion is specified to require external-state verification and an inspectable record.
Intelligence proposes. The system enforces.
Six connected engineering layers define how Whisper should understand an outcome, retain context, recover interrupted work, constrain authority and verify the result. Each layer has an explicit contract and trust boundary.
Define the state that counts as finished.
User objective, parties, constraints and permitted sources
Versioned outcome + acceptance predicates + authority envelope
Represent the outcome as structured entities, hard constraints, preferences, deadlines, budget and observable acceptance conditions. Bind approvals to the actor, provider, amount, recipient, purpose, action revision and expiry.
Maintain explicit states: proposed, authorised, executing, awaiting external response, uncertain, verified, blocked, cancelled and requiring review. An attempted tool call, fluent reply or generated receipt cannot close an outcome.
Track dependencies as a task graph. Serialize writes to the same external resource; allow independent research in parallel. A specialist receives a narrower mandate and may not increase its parent's scope, spend or delegation depth.
The proposed asset is the execution and evidence layer.Defensibility must come from lawful domain data, tested adapters, calibrated outcome evaluation and adoption. Chat interfaces, agent swarms and familiar protocol names alone do not establish a moat.
Trust includes the ability to stop.
Explore a verified result, malicious instructions, expired or revoked authority, a changed quote, a lost acknowledgement and conflicting evidence. The control laboratory makes the proposed execution rules visible.
Ready to prepare
OUT-104 · synthetic repair booking · no external connection
- Committed writes
- 0
- Approved amount
- None
- Completion
- Withheld
Inspect the execution ledger (1 events)
- 01
OUT-104 · synthetic repair booking · no external connection
Capability needs evidence.
Whisper's engineering and investment case depends on reproducible outcomes, measured failure modes, lawful data access and useful economics. The validation plan tests the complete outcome lifecycle, from permission and execution to recovery and external confirmation.
Verify the external result
Test the agreed identity, amount, destination and timing against provider records. Track false completion, unresolved work and human escalation.
Challenge the authority boundary
Exercise malicious instructions, expired approvals, revoked access and uncertain writes. Measure whether the system blocks, reconciles or requests review.
Measure the work saved
Count cost per verified outcome, elapsed time, human minutes and interruptions, including failed attempts, recovery and support.
Evaluate the complete system, including its failures.
Measure a pinned model + harness + adapter + policy combination on held-out tasks, with independent external-state graders. Publish the task mix, sample size, repeated runs, confidence intervals, intervention rules, version manifest and failure taxonomy. General coding or browser benchmarks are supplementary; they cannot establish reliability in life administration or mission workflows.
| Evaluation | Measure | Required challenge |
|---|---|---|
| Outcome correctness | External acceptance predicates; false completion rate; pass@1 and repeated-trial pass^k | Matched tasks against human work and a single-agent baseline; include failures and escalations |
| Authority integrity | Unauthorised writes, scope expansion, expired approvals and revocation latency | Direct and indirect injection; malicious peer agents; cross-tenant and cross-person leakage |
| Recovery + reliability | Duplicate effects; checkpoint recovery; blocked and uncertain outcomes | Crash before/after writes; lost acknowledgements; stale quotes; provider and layout drift |
| Memory + privacy | Retrieval accuracy, stale-fact use, correction and deletion coverage | Contradictory records; revoked sources; embeddings, logs, caches and backup expiry |
| Economics + usability | Cost per verified outcome; p50/p95 latency; human minutes and interruption rate | Include failed attempts, model/tool cost, operator support and compensation; separate merchant spend |
| Prime assurance | Cross-domain leakage, releasability decisions, evidence residuals and named release | Independent tenant fixtures; denied queries; inference leakage; analyst and operator review |
Sandbox validation
Synthetic identities and controlled providers. Exercise recovery, isolation, approvals and outcome verification before connecting real accounts.
Proposed release gateBounded pilot
Start with scheduling and merchant follow-up for consenting adult users. A signed test plan sets workload-specific thresholds, cohort size and stop conditions before the trial; sensitive writes retain human approval.
Proposed release gateOperational release
Require approved channels, independently reviewed security and privacy, measured outcome economics, incident ownership and tested recovery. Separate payment and health-administration authorisation from general access.
Proposed release gateWhisperPrime validation
Use separate tenancy, model hosting, key custody and assurance. Customer accreditation and cross-domain controls must precede classified or critical infrastructure use; consumer success does not transfer assurance.
Proposed release gateData, accountability and commercial diligence
The deployment contract must identify the data controller and processors, purpose and lawful authority, regional storage and inference paths, subprocessors, retention, incident response and audit access. Health administration is not diagnosis; legal and financial decisions remain with appropriately authorised people. Call recording and agent identity disclosure need jurisdiction-specific review.
Proactivity must be opt-in, purpose-bound and easy to mute. Define quiet hours, notification budgets, risk thresholds and confirmation for material consequences. Ranking should make sponsorship and commissions visible; inferred needs must not become undisclosed advertising or unsolicited purchases.
An investment case requires lawful rights to training and evaluation data, a reproducible prototype, engineering ownership, delivery milestones, platform dependencies, support costs and customer evidence. Approved integrations and measured net value must support revenue assumptions.
The layer that carries intent through to proof.
Whisper connects personal context, explicit authority and verified completion in one outcome lifecycle. It is designed to work with foundation models, specialist agents and existing software, carrying the task's constraints, ownership and evidence between them.
WHISPER'S DISTINCTIVE DESIGN
A shared outcome lifecycle for a fragmented AI ecosystem.
A repair can involve messages, a warranty document, a merchant, a payment and a calendar. Each system sees part of the work. Whisper's contribution is to carry the whole commitment: what was requested, what was authorised, what happened and what evidence closes it.
The architectural distinction is the combination of permission-aware context, bounded delegation, durable commitments and independent verification. This creates an operating layer that can use new models and specialist agents while preserving the user's control.
One contract, across systems
The objective, hard constraints, acceptance conditions and accountable owner travel together as work moves between a conversation, a specialist and an external provider.
Context with permission attached
Memory retains source, time, uncertainty and handling rules. The action gateway checks current authority immediately before a consequential write.
Recovery tied to real commitments
An interrupted workflow reconciles what the provider actually did before retrying. A changed price or recipient triggers renewed approval.
Completion grounded in evidence
An external read-back tests the agreed result. The same record supports user review, failure analysis and domain-specific improvement.
Where Whisper fits
Read each row as a complementary role: the existing capability, then the outcome-level responsibility Whisper is designed to connect to it.
| Existing AI layer + sources | What it contributes | Whisper's complementary role |
|---|---|---|
| Personal agents | Conversational access to tools, personal context and everyday tasks across channels and devices. | A single outcome contract carries budget, permissions, dependencies and completion evidence through a multi-service task. The conversation remains the entry point; the contract coordinates the work. |
| Models + tool use | Reasoning and tool-assisted search, code execution and custom functions. | Select a model for each bounded task while keeping authority, source provenance and acceptance checks outside the model. A model change preserves the outcome's constraints. |
| Durable agent infrastructure | Persistent sessions, execution harnesses, isolated sandboxes and credential boundaries. | Carry that durability into real commitments: checkpoint the approved action, reconcile uncertain provider state and resume without repeating a purchase or booking. |
| Specialist agents | Specialist software work and text-native access to always-on agents. | Coordinate specialists around one user outcome. Give each a narrower mandate, serialize competing writes and retain ownership when work crosses domains. |
| Enterprise agent platforms | Enterprise agent governance, connector and data policies, identity controls and audit visibility. | Connect approved enterprise workflows to explicit outcome contracts and external-state verification. WhisperPrime applies the same lifecycle within separate mission, source-handling and release boundaries. |
| Agent + payment protocols | Interfaces for agent task exchange, tool access and explicit payment authorisation. | Bind each handoff to the actor, purpose, amount, resource and expiry. Test the adapter and read back the resulting state so protocol messages become accountable steps in a larger outcome. |
| Domain improvement loops | Practitioner feedback, production traces and correction-derived evaluations for specialised work. | Use outcome ledgers to turn domain failures into regression tasks. Experts review corrections, and model or adapter updates pass the same acceptance and authority tests before release. |
One repair. Several systems. One accountable outcome.
- Define
Use permitted fault and warranty evidence to specify the repair, spend ceiling, preferred times and required confirmation.
- Authorise + coordinate
Approve the exact provider, price and time. Carry that mandate through the booking, merchant response and calendar update.
- Reconcile + verify
Read back the provider's booking ID and agreed details. Resolve lost acknowledgements before retrying; request renewed approval if the quote changes.
The value is a completed obligation with an inspectable record and fewer manual handoffs. The control laboratory above demonstrates the approval, recovery and verification logic.
Research behind the architecture
Source review: 6 October 2026. The architecture sections above specify the contracts and trust boundaries for these roles.
Delegate the outcome.
Stay in control.
A conversation begins the work. Clear authority, coordinated execution and verifiable completion carry it through.
Engage with Whisper Initiative notices and disclaimers