R&D / Alakris Lab

Alakris Lab

Researching Autonomous AI Organizations

Alakris Lab is the research and development initiative of Alakris focused on autonomous AI organizations — systems in which AI agents perform real operational work, coordinate with people and other agents, make decisions, recover from failures and operate with increasing autonomy.

Alakris is building and researching autonomous AI organizations — systems where AI agents evolve from assistants into digital employees and coordinated AI teams capable of performing real organizational work.

Explore our research Collaborate with Alakris Lab

AI Agents
AI Employees
AI Teams
Autonomous AI Organizations

Alakris Lab

What is Alakris Lab?

Alakris Lab is the research direction of Alakris. Real commercial deployments of AI agents provide a practical environment for studying not only model quality, but the reliability of long-running organizational work.

How can AI systems reliably perform long-running organizational work in real-world environments?

Our Research Thesis

Alakris studies the architectures, methods and metrics required for specialized agents to share work, preserve authority boundaries, recover from failure and remain accountable to people.

We believe the next major step in applied artificial intelligence is not a single autonomous agent, but autonomous organizations composed of multiple specialized AI agents operating together with humans.

R&D

Research Agenda

Eight connected research areas frame the long-term Alakris Lab programme.

Multi-Agent Coordination

How AI agents divide tasks, transfer context and coordinate work.

Long-Horizon Autonomy

How agents execute work lasting hours, days or weeks without continuous supervision.

AI Employee Reliability

Reliability of digital employees, error detection and recovery after failures.

Human-Agent Collaboration

How people delegate work and control the level of agent autonomy.

Agent Memory & Context

Long-term memory and context preservation across tasks.

AI Governance

Permissions, authority, policies and limits for autonomous agents.

Economics of AI Work

Task cost, productivity and careful comparison of AI and human work.

Autonomous Organization Benchmarks

Methods for measuring autonomy, quality and efficiency of AI organizations.

Research Novelty

Areas where Alakris aims to contribute new knowledge include:

Rather than studying agents only in simulated environments, Alakris aims to evaluate autonomous systems through real operational deployments.

Measurement of organizational AI autonomy
Benchmarks for real-world AI work
Coordination architectures for multiple AI employees
Metrics for human intervention
Recovery mechanisms for long-running agents
Governance models for autonomous AI teams
Economic models of AI-operated organizations

Research Through Real-World Deployment

Real products create a measurable research loop while private customer data remains outside the public research surface.

Deployment
Telemetry
Dataset
Experiment
Benchmark
Research
Better Product

Task completion rate
Human intervention rate
Cost per task
Latency
Failure rate
Recovery rate
Autonomy level
Long-horizon task completion

Extreme Autonomy

Extreme environments are a useful research framework for studying autonomy under limited communication, constrained resources, system failures and delayed human intervention.

We see extreme environments as a long-term stress test for the architectures and metrics developed for autonomous AI organizations — not as a claim of a current space programme or partnership.

Remote industrial environments
Space operations
Planetary exploration
Future Mars missions

Research Collaboration

Alakris Lab is open to dialogue with researchers, universities, AI laboratories and technology companies working on autonomous agents, multi-agent systems, AI safety, robotics and organizational AI.

Discuss Research Collaboration

Joint experiments
Benchmark development
Dataset research
Technical reports
Academic publications
Applied research pilots
Research internships
University collaborations
Industrial PhD programmes

Publications & Research

Published materials and ongoing research by Alakris Labs as of 6 October 2026. Preprints, manuscripts, preregistration and data have distinct statuses; they are not all peer-reviewed publications.

Papers

Technical Reports

Benchmarks

Datasets

Open Source

Research programme

Digital Organizations Beyond Human Reach

A long-term programme on autonomous organizations when human authority is unreachable, persistent posts and institutional memory. Future topics include organizational recovery, distribution of authority and mission continuity. This is a research agenda, not a promise of available products.

View source

Preprint · 20 Aug 2026

Escalation Presupposes a Reachable Addressee

A position paper on the reachable-addressee assumption behind escalation and structurally independent oversight. Published on Zenodo as a preprint, DOI 10.5281/zenodo.22030089.

View source

Research manuscript

Structural independence of an onboard oversight role under communication delay

An empirical manuscript from a preregistered simulation. The independence-by-delay interaction was supported; equivalence at zero delay remained inconclusive. A second-model check exposed a cost of independence when escalation is available. Paper acceptance is not confirmed.

View source

Preregistration · 2 Sep 2026

Organizational Independence of Oversight Roles in Communication-Delayed Autonomous Missions

A factorial-study plan: six oversight configurations, two delay conditions and 25 confirmatory seeds. OSF registration separates prespecified hypotheses from subsequent analysis.

View source

Open data · 3 Sep 2026

egsim — pilot materials

Pilot data, calibration reruns and frozen code. The pilot informs design and power estimation; it is not confirmatory evidence.

View source

Open data · 15 Sep 2026

egsim — confirmatory materials

270 experimental cells, frozen analysis, code and materials with a reproducible execution procedure. This deposit concerns simulation; real-mission generalization requires separate testing.

View source

Open source

egsim — escalation-gap simulator

Simulator source, scenarios, experimental harness, research texts and reproduction instructions. The public repository makes the method and its limitations inspectable.

View source

Scientific community, workshops and working groups

Confirmed activities and open discussions. Reviewing, submitting a paper and giving a talk are different forms of participation; no paper acceptance or completed presentation is claimed here.

Discussion · since September 2026

AAIF Identity & Trust: authority when escalation is unreachable

We introduced an authorization use case for an unreachable external authority. Discussions cover exact permission scope, revocation and fail-closed behavior. Joint tests of binding vetoes, stale evidence and already-committed effects are proposed; results are not yet available.

View source

Open use case · October 2026

Joint comparison of decision, effect and outcome

Issue #13 proposes a reproducible runner to compare Alakris, MintID and Proofable approaches: authority at dispatch, committed effects and task outcomes are recorded separately. The shared experiment is in preparation.

View source

Research contacts · 6 Oct 2026

AAIF Identity & Trust participant map

A sourced selection of ten participants in our discussions, their roles and project context. This is a contact map, not an official AAIF directory or a list of Alakris partners.

View source

Invitation accepted · 17 Sep 2026

NeurIPS 2026 Meta-Agents: reviewer service

Vladislav Kostitsyn accepted an invitation to review for Managing Agents that Manage Agents. This is workshop reviewer service, not main-conference reviewing, an accepted paper or a speaker role.

View source

Research Network

We contribute to AAIF Identity & Trust discussions and compare approaches to agent authority. Participation by individual researchers does not imply an Alakris partnership with their employers. The community section links to contacts and public sources.