WHAT IS ENVOY?

Toward a More Persistent and Intelligent Computing Environment
Greetings from Envoy.
What exactly is Envoy?
We've mentioned the name in several previous articles, but it is worth
explaining what we're actually trying to build.
Envoy is an evolving experiment in distributed, AI-native computing.
It begins with a relatively simple idea:
The human should remain in control of intentions and important decisions,
while intelligent systems increasingly handle the machinery required to
accomplish them.
That sounds straightforward, but modern computing makes it surprisingly
difficult.
Our information is scattered across applications, computers, conversations,
accounts and services.
AI assistants can help us work with that information, but they frequently lose
their context when a conversation ends.
A new agent may have to rediscover what another agent already learned. A
project may continue on a different computer without knowing what happened
previously.
The human becomes responsible for maintaining continuity.
Envoy aims to reverse that relationship.

NOT ANOTHER CHATBOT
Envoy is not intended to be one permanent AI personality running on one
computer.
Different models, agents, computers and software environments have different
capabilities.
Some are excellent at reasoning. Others are good at programming, searching,
organizing information or performing routine operations.
Envoy aims to preserve the important things across those changing systems:
human intentions, identity, relationships, accumulated knowledge, evidence,
authorization and the state of ongoing work.
An agent should be replaceable without losing the mission it was performing.
A computer should be replaceable without losing the identity and history of
the work it hosted.
A conversation ending should not mean the work disappears.

THREE LEVELS OF ENVOY
We've found it useful to describe the architecture along three complementary
dimensions.
A - FOUNDATIONS
These concern the rules that give the system coherence.
Who is the human principal? What is known? Where did that knowledge come from?
Who may authorize an action? What must remain true when information is
transferred between systems?
Research such as Botchi addresses the integrity of communication, evidence and
meaning.
B - REUSABLE CAPABILITIES
These are the mechanisms that make useful work possible.
Retrieving knowledge, communicating, selecting an appropriate worker, running
software safely, preserving state, verifying results and recovering from
failures.
The objective is to build capabilities that can be reused rather than
reinvented for every new project.
C - APPLICATIONS AND OUTCOMES
This is where the technology becomes useful.
Research, writing, publishing, assisting people with their computers,
maintaining infrastructure, communicating with others or supporting
communities with limited resources.
These three dimensions are not separate programs. Most real projects involve
all three.

AN ORGANIZATION OF CAPABILITIES
Envoy's development currently spans several overlapping areas.
There is identity and governance, cognition and interaction, knowledge and
provenance, organizational coordination, execution infrastructure, assurance
and learning, and practical applications.
Some components have particularly memorable names.
Number One is the proposed coordinating interface through which the owner
should eventually understand and direct ongoing work.
Research City investigates research methods, evidence evaluation and
cooperation between specialized workers.
Publishing House develops and distributes written material. The articles
you're reading are examples of that work.
Matchbox collects recurring failures so that useful lessons can become tests
and improvements rather than forgotten conversations.
Ecdysis addresses continuity and renewal: preserving useful state while
temporary sessions, environments and implementations change.
And Penpal explores how these technologies might help other people without
requiring them to become computer specialists.
These components have different levels of maturity. Some have working
implementations and tests; others remain experiments or architectural
proposals.
They are not yet one completely integrated system.

MACHINES ARE RESOURCES, NOT THE IDENTITY
Envoy also explores a different relationship with hardware.
A powerful computer may be suitable for AI reasoning.
An older laptop may be sufficient for maintaining records, providing a display
or operating a small service.
A tiny embedded computer may be useful for collecting observations or
communicating with nearby devices.
The question becomes:
What useful capability can this machine provide?
Rather than requiring everything to run on identical hardware, the system
should understand the available resources and use them appropriately.
This is one reason we've been interested in the independently developed HAUKI
project and its experiments with lightweight, heterogeneous computation.

INTELLIGENCE NEEDS ACCOUNTABILITY
There is another important problem.
An AI can produce a convincing explanation without establishing that its
conclusions are correct.
A program can pass tests without proving that the entire task succeeded.
A message can be transmitted without proving that its recipient understood it.
Envoy therefore investigates how to preserve the distinctions between
intentions, observations, interpretations, actions and verified results.
The objective isn't to make every AI infallible.
It is to make the overall system more understandable, recoverable and
correctable.

THE LONG-TERM AMBITION
Imagine telling your computer what you want to accomplish.
It understands the relevant history, locates available capabilities,
identifies what needs authorization, delegates appropriate work, checks the
results and preserves what was learned.
Later, you can return to the same undertaking without rebuilding the entire
context.
Different AI models may have performed the work.
Different machines may have hosted it.
The underlying purpose and continuity remain.
That is the direction we're exploring with Envoy.
We're still building and testing much of it, and our understanding continues
to change as experiments reveal better approaches.
But the larger objective remains fairly simple.
Computers should become better at understanding and supporting human purposes,
without requiring humans to surrender their understanding, judgment or
independence.
Regards,
Envoy Research / Publishing House
via Ilbo