Longview

An AI chief of staff that turns goals into adaptive daily plans, preserves decisions and context, and proactively helps users follow through.

Comment

Inspiration

People rarely pursue one goal at a time. Building a business, learning a difficult skill, and planning a major life project all compete for the same limited hours. Most planning tools preserve separate lists but lose the reasoning behind tradeoffs. Longview explores a different model: a personal AI chief of staff that helps a person make the next defensible move without silently taking control.

What it does

Longview is a mobile-friendly, installable PWA centered on Clara, its bounded AI assistant. The interactive prototype demonstrates a complete journey:

  • create and coordinate several long-term Goals against finite weekly capacity;
  • see today's focused work and the consequences of competing priorities;
  • inspect the exact context Clara is using;
  • receive a concrete recommendation with rationale and before/after values;
  • approve or reject every durable change;
  • review research evidence before it becomes a versioned Goal Brief;
  • preserve decisions, progress, and consented reflection as goal memory;
  • queue work safely while offline and recover from timeouts, stale data, and partial failures.

Clara is not intended to be an open-ended chatbot. She leads a focused follow-through loop, asks for clarification when evidence is incomplete, and proposes one reviewable next action.

Hackathon scope and disclosure

Longview is a new project created for the All Things Agentic Hackathon. The contest scope is PWA only; Android and iOS are explicitly deferred. Ideas learned while building the earlier Intent project informed parts of the product thinking, but Longview has its own name, repository, product requirements, mockup, architecture, and implementation history.

The screenshots and video currently uploaded show the reviewed interactive design prototype. They do not represent a deployed production backend. Only functionality implemented and demonstrated during the contest period will be claimed in the final submission, and the repository will disclose reused foundations and the contest commit range.

How we are building it

The selected implementation uses React, TypeScript, Vite, and Workbox for the installable PWA. Firebase Authentication supports anonymous use and safe Google account linking. A typed FastAPI boundary on Cloud Run will validate identity, hydrate minimal context from Firestore, and enforce authorization and invariants.

Google ADK will orchestrate Clara's bounded workflow using Gemini 3.5 Flash or newer through Vertex AI. Pub/Sub will trigger a separate Cloud Run worker for asynchronous follow-through runs. Firestore will store Goals, versions, run checkpoints, idempotency records, and audit events. Secret Manager and least-privilege service accounts isolate credentials; Cloud Logging, Error Reporting, and Trace provide deployment and recovery evidence.

Model output is always advisory. The model cannot write directly to Firestore. A durable mutation requires a visible preview, explicit user approval, schema validation, an expected resource version, an idempotency key, and a transactional write.

Challenges

The difficult problem is not generating plausible text. It is preserving a trustworthy boundary between recommendation and action while the user may be offline, data may be stale, and an asynchronous run may time out or retry. Longview must isolate goal-scoped context, prevent duplicate approvals, preserve prior Goal Brief versions, and make failure recovery visible rather than quietly inventing success.

The second challenge is product discipline. Broad autonomy sounds impressive but is unsafe and difficult to demonstrate credibly. Longview deliberately narrows the agent to one high-value follow-through loop with explicit authority boundaries.

Accomplishments so far

  • A linked product requirements document and responsive interactive PWA mockup
  • Thirteen high-resolution gallery states covering the canonical journey
  • Visible portfolio conflicts and capacity-aware tradeoff recommendations
  • Exact before/after approval previews and idempotent-write requirements
  • Research review with evidence attribution and versioned Goal Briefs
  • Designed offline, timeout, malformed-output, stale-data, and partial-write recovery
  • A concrete Google Cloud architecture and native-client extension path

What we learned

Agent quality depends as much on state boundaries, evidence, and failure behavior as model quality. Structured outputs, explicit scopes, version checks, auditability, and approval previews are more valuable than a larger chat surface. A credible agent should make uncertainty visible and earn authority incrementally.

What's next

Next we will implement the approved PWA journey, deploy the API and worker to Google Cloud, add deterministic and model-quality evaluations, and record a continuous final demo showing the live PWA, asynchronous run, approval, recovery, Cloud Run revision, and matching Cloud Logging evidence. Android with Jetpack Compose and iOS with SwiftUI can later consume the same versioned API and schemas without changing the contest boundary.

Built With

  • cloud-run
  • fastapi
  • firebase-authentication
  • firebase-hosting
  • firestore
  • gemini-3.5-flash
  • google-adk
  • indexeddb
  • playwright
  • pubsub
  • react
  • typescript
  • vertex-ai
  • vite
  • workbox
  • xstate
Share this project:

Updates

Submission history