Seven pillars between today's capable models and reliable, collaborative Intelligent Machine Systems. Every pillar links to the public artifacts behind it. Last updated October 2026.
Evidence first
Every pillar links to public artifacts: source, live systems, or documented architecture. If there is no artifact, the pillar does not claim a result.
Built in the open
The research program runs in public repositories. You can read the code, the review design, and the history instead of taking a claim on faith.
Honest about limits
The claim ladder below states exactly how far the self-improvement work has gone and where a person still approves every change. Nothing beyond that is claimed.
The question each pillar answers, why it matters to an operator, and where the work stands today.
Reasoning that stays correct as the world and interface shift underneath the agent.
The question
How does an agent keep its decisions valid when the situation changes faster than the model was trained for?
Why it matters
Operational feeds do not pause for a model. A system that reasons over stale state produces confident, wrong answers at the worst moment.
Decisions anchored to observable state, not just internal priors.
The question
How does an agent verify what is actually true right now instead of pattern-matching what was true in training?
Why it matters
A recommendation nobody can trace to an observable source cannot be signed for. Grounding is what makes output reviewable.
Understanding across text, images, screens, and video - including implicit visual structure.
The question
How does an agent read a screen, a map, or a video feed the way an experienced operator does, including what is implied rather than labeled?
Why it matters
Most operational software was built for human eyes. An agent that cannot read what is on the screen cannot do the work that matters.
Memory that compounds, paired with the ability to judge and improve its own work.
The question
How does an agent carry what it learned yesterday into today, and how does it catch its own mistakes before a person has to?
Why it matters
An agent that starts from zero every session never gets better at your work. One that cannot grade itself needs a person grading it forever.
Representing real software actions as learnable, verifiable units of work.
The question
Can a unit of digital work be defined precisely enough that a machine can learn it, a buyer can price it, and an auditor can verify it?
Why it matters
Machine labor cannot be bought, managed, or audited until the unit of work is explicit. This pillar is the accounting layer for everything else.
Ecosystems where agents and tools cooperate through open protocols.
The question
How do independent agents hand work to each other, with context intact, without a bespoke integration for every pair?
Why it matters
Single agents plateau. Operational value comes from fleets, and fleets need a shared protocol instead of point-to-point glue.
Measuring, auditing, and bounding agent behavior for trust at scale.
The question
What does an agent need to record about itself so that a person can approve, audit, and roll back its work without watching every step?
Why it matters
This is the difference between a demo and a deployment. Every other pillar ships through this one or it does not ship at all.
The ladder is an honest map of what our systems credibly do today, from tooling that tidies itself up to a full intelligence explosion. Levels one through three are standing capability within bounds. Levels four through six are partial, each with a person approving the step that matters. Level seven is claimed by no one, including us.
Recursive Self-Improvement Claim Ladder
Claim
AGI-level intelligence explosion
Your current status
No public system has conclusively demonstrated this
Claim
Recursive improvement of the research loop
Your current status
Partial: the system adjusts its own research approach. Any change to how that works is approved by a person first
Claim
Automated AI-research loop
Your current status
Partial: an automated research loop over models we already control. Actual training runs still need a person's go-ahead
Claim
The main model adapts how it works
Your current status
Partial: the system tunes how it behaves within limits set in advance. Its underlying model is untouched
Claim
It trains smaller, focused models
Your current status
Yes, within bounds: it trains smaller models locally, and only when the conditions we agreed beforehand are met
Claim
Code and policy changes, approved by a person
Your current status
Yes, within bounds: every change is checked, approved by a person, and can be undone
Claim
The tooling around the model improves itself
Your current status
Yes, credible
Short, factual notes on research and deployments as they happen.
As of October 2026, AI capability is accelerating across coding, computer use, multimodal reasoning, scientific work, and Intelligent Machine Systems (IMS) - but reliability, autonomy boundaries, safety, and evaluation remain active challenges. ELI Labz treats these challenges as the core of the work, not a footnote.
Put this page in front of the right person
This page travels by forwarding, not by advertising. If it describes someone's week, send it to them directly.
Bring one workflow that cannot stay a demo. ELI Labz embeds with your Crew, builds the Intelligent Machine System around it, and proves it with records you can check.
Advisory from $250 per hour (2-hour minimum) • Deployment sprints from $18,000 • Embedded from $32,000 per month • Retainer from $16,000 per month
Prefer email? dmesa@e-l-i.net