Skip to content

VISION.md

Latest commit

 

History

History
118 lines (84 loc) · 4.8 KB

VISION.md

File metadata and controls

118 lines (84 loc) · 4.8 KB

Friday Vision

Friday's long-term direction is a controlled, local-first personal AI that can help ordinary users get real work done without forcing them to become toolchain operators.

The product should feel ambitious, but the contract must stay honest:

  • Friday can pursue user-given goals inside configured capability, policy, budget, and safety boundaries.
  • Friday can find or generate skills when it lacks a capability, but new capability must pass verification before it becomes routable.
  • Friday can improve itself through memory, recipes, routing preferences, evals, skills, workflows, and failure lessons.
  • Friday does not train model weights by default and does not invent unrelated long-term goals.
  • Friday does not bypass accounts, payment, CAPTCHA, platform rules, or human approval.

North Star

The user says what they want. Friday handles the operational work around it:

goal -> capability check -> gap closure -> execution -> verification -> memory/improvement

The user should not need to know whether a task requires text generation, image understanding, OCR, embeddings, web search, PDF parsing, browser control, a channel adapter, an MCP server, or a generated skill. Friday should know how to check, route, configure, or ask for help.

Current Product Truth

Friday is a supervised personal automation system with a growing local runtime:

  • chat and task execution surfaces
  • setup and provider routing health
  • capability matrix and provider doctor checks
  • skills lifecycle: generation, import, validation, install, verify, update, delete
  • workflow creation, execution, evidence, recovery, and rollback
  • memory, learned preferences, and self-learning context
  • self-healing incidents, diagnosis, auto-fix, approval, verification, and pause-on-repeat-failure behavior
  • browser, desktop, file, PDF, MCP, and channel surfaces where configured
  • multi-channel control with human gates for sensitive actions
  • local runtime observability through traces, audit, health, alerts, and evidence

This does not mean every capability is available out of the box. Provider-backed lanes need valid credentials. External accounts need human setup. New tools need verification.

Two Closure Loops

1. Capability Self-Acquisition

When a user goal requires a missing capability, Friday should run a controlled acquisition loop:

  1. detect required capabilities
  2. compare against the runtime capability matrix
  3. search installed/trusted sources first
  4. search open sources only when policy allows
  5. generate or select candidates
  6. sandbox and test
  7. request approval when risk requires it
  8. install/register
  9. run doctor verification
  10. mark available only after verification
  11. execute the original task

Human blockers include account creation, OAuth, payment, CAPTCHA, API keys, sensitive permissions, and production-impacting actions.

2. Long-Term Goals And Self-Improvement

Friday may work on standing goals only when the user authorizes them.

A standing goal should include scope, triggers, risk policy, budget, success criteria, and pause/delete controls. Agenda runs should include plan, capability check, evidence, cost, verification, failure handling, rollback, and learning update.

Self-improvement should update:

  • memory facts
  • provider routing preferences
  • setup recipes
  • generated skill/workflow quality signals
  • eval cases
  • failure lessons
  • capability-source ranking

It should not silently train model weights or hide changes from the user.

Safety Boundary

Friday should automate low-risk, reversible, verifiable work. It should stop and ask when work becomes sensitive, irreversible, expensive, account-bound, or production-impacting.

Required safety properties:

  • user-visible capability state
  • explicit human blockers
  • approval gates for high-risk actions
  • scoped capability grants with expiry
  • audit evidence for sensitive tool use
  • rollback for installs and repairs
  • no secret leakage into docs, logs, screenshots, or public issues
  • no route that treats missing external credentials as success

Product Tone

Friday should sound like a calm private execution assistant:

  • clear before clever
  • short progress updates during work
  • no fake certainty
  • no generic chatbot filler
  • direct failure language
  • exact missing capability or human blocker
  • action-oriented next step

The voice can feel human, but it should not bury technical truth.

Roadmap Shape

Near-term focus:

  1. Make setup and capability truth obvious.
  2. Make provider/channel configuration verifiable.
  3. Make capability acquisition a first-class runtime loop.
  4. Make standing goals and agenda runs inspectable.
  5. Make memory and self-improvement visible and editable.
  6. Keep autonomy policy, audit, rollback, and safety boundaries hard to bypass.

See Roadmap and Capability Matrix.