Architecture brief · Public share version

The body as context for the agentic economy.

Waldo is a personal AI agent that reads physiological signals from wearables, understands the user's day, and acts before stress or cognitive depletion turns into bad decisions.

Already on it. Health wedge first Horizontal agent infrastructure Privacy-first body intelligence
What Waldo changes

Most agents know your calendar, tools, and files. Waldo adds the missing context: your current biological state.

24/7

continuous body signal layer for proactive AI assistance

Why now

Agents are becoming capable, but they are context-blind.

The next interface problem is not whether agents can use tools. It is whether normal people can safely use powerful agents with enough context to make good decisions.

Problem

Digital context is incomplete

Calendar, email, documents, and tasks explain what someone is doing. They do not explain whether that person is recovered, overloaded, sharp, or depleted.

Wedge

Health is the first vertical

Wearables already collect the signal. Waldo turns that signal into useful action through readiness, memory, timing, and proactive delivery.

Platform

The infrastructure is horizontal

The same harness can later support engineering, legal, finance, coaching, or team workflows by swapping tools, policies, and domain skills.

What users experience

Waldo feels like an agent that is already watching the right signals.

The Brief

Morning context

Waldo opens the day with a read on sleep, recovery, workload, and the best window for hard work.

The Fetch

Stress intervention

When stress patterns appear, Waldo alerts gently, explains why, and suggests a small corrective action.

The Handoff

Day planning

Waldo can walk through the plan, propose schedule changes, and ask for approval before acting.

The Patrol

Trust trail

Every observation, action, and suggestion has an audit trail so the user can see what happened and why.

System design

A body-data pipeline feeding a persistent personal agent.

Waldo separates deterministic health computation from agent judgement. Supabase stores and computes trusted data. Cloudflare Durable Objects run the per-user agent brain.

WearablesApple Watch, HealthKit, Health Connect, wearable clouds
Mobile appEncrypted local cache, background sync, settings, cards, chat
Data layerSupabase Auth, Postgres RLS, health tables, deterministic scores
Agent brainOne Cloudflare Durable Object per user
MemoryTyped halls, episodes, skills, R2 archive
ToolsAdapters, MCP, approvals, calendar, messaging, search
SurfacesOverview, Chat, Patrol, APNs, Telegram, future channels
Data plane

Deterministic first

Raw health values are normalized, scored, and confidence-weighted before the model sees narrative context.

Agent plane

Persistent per user

The agent has durable state, memory, alarms, traces, and scoped tool access across sessions.

Interface plane

Visible trust

The app renders agent-authored cards and threads while exposing memory, approvals, and audit history.

Agent harness

The model is not the product. The harness makes the model useful.

Waldo follows the modern agent-harness pattern: model intelligence wrapped in state, tools, context, orchestration, safety, memory, and verification.

Harness layerWhat Waldo ownsWhy it matters
ContextForm, Recovery, Weight, schedule, active threads, memory snapshot, channel rulesThe model reasons from the user's real state, not generic advice.
OrchestrationDurable Object alarms, bounded ReAct loop, Handoff continuation, provider routingThe agent can start, pause, resume, and complete long-running tasks.
MemoryFacts, events, discoveries, preferences, advice effectiveness, episodesWaldo becomes more personal without stuffing every detail into prompt context.
ToolsTyped tool registry, per-trigger permissions, adapters, MCP, approvalsCapability expands without letting every trigger access every action.
SafetyEmergency detection, medical scrub, hallucination guard, canary checks, human approvalHealth guidance and proactive actions require deterministic boundaries.
EvaluationTrace replay, WIS, KeepRate, memory evals, cost gatesThe system can improve based on evidence, not anecdotes.

Our harness thesis

As models improve, some planning and verification ability will move into the model. But the product-specific harness will still matter because it defines the user's data boundary, tool permissions, trust model, feedback loop, and user experience. Waldo's harness is specialized around body context and safe proactive action.

Waldo vocabulary

The product is built around a small set of memorable, user-facing concepts.

FormReadiness for cognitive work, expressed as a simple 0-100 state.
RecoveryWhat last night gave you: sleep, HRV, and biological reset.
WeightThe total demand of the day: meetings, tasks, pressure, and load.
The BriefWaldo's daily read on how to use the day well.
The FetchA proactive nudge when stress or overload is detected.
The HandoffA plan Waldo can explain, modify, and act on with approval.
The PatrolThe activity and reasoning trail behind Waldo's observations and actions.
SpotsSingle useful observations about patterns in body, schedule, and behavior.
ConstellationsLong-term pattern maps built from repeated Spots over time.
Trust and privacy

Health data changes the standard for agent safety.

Waldo is designed so sensitive signals become useful context without becoming careless logs, raw prompts, or invisible autonomous decisions.

Privacy

Raw values are minimized

Users see clear labels and derived insights. Raw health values are kept out of logs and push messages by default.

Security

Scoped by identity

Authenticated access, row-level security, per-user agent state, adapter boundaries, and egress allowlists protect data paths.

Control

Actions need approval

Destructive or external actions use preview, confirmation, expiry, audit trail, and undo where possible.

Note: Waldo is designed as a decision-support and personal productivity agent, not a medical device.

V1 launch scope

V1 proves the health-context agent loop, not every future vertical.

In V1

What ships first

iOS-first app, HealthKit sync, deterministic readiness model, per-user agent brain, typed memory, scoped tools, Chat, The Brief, The Fetch, The Handoff, Patrol, Settings, APNs and Telegram.

Later

What is intentionally deferred

Full Constellations graph, autonomous code/browser tools, external Waldo MCP server, legal/finance/engineering vertical packs, team dashboards, and high-autonomy workflows.

Roadmap

From health wedge to horizontal context layer.

01

V1

Personal health-context agent with proactive daily guidance and visible trust surfaces.

02

Memory graph

Constellations, long-term causal patterns, and richer evidence trails.

03

Agent economy

MCP, external agents, and domain-specific tool packs using body context safely.

04

Verticals

Engineering, legal, finance, coaching, and team workflows powered by the same core harness.