Identify the system.
Model, version, agent roles, tools, permissions, environment and time of observation belong with every performance claim.
INDEPENDENT AGENT SYSTEMS RESEARCH / 001 — OPEN EDITION
Autonomy
needs evidence.
How AI agents plan, coordinate, use tools, and make decisions—and how we determine what they actually do, what can fail, and who stays in control.
01 / AGENT ARCHITECTURES02 / REAL-WORLD EVALUATION03 / HUMAN CONTROL
AGENT↗
PLANTOOLVERIFYOUR POSITION
We do not accept demonstrations as proof of reliability. We examine systems, permissions, measurements and failure modes with the same scrutiny as their successes.
THE RESEARCH / TEN FIELDSFollow the
Follow the
system boundaries.
Ten distinct lines of inquiry. Each dossier establishes its scope, source trail, unanswered questions and conditions for future evaluation.
01 / RESEARCH AREA
Agent Foundations
What actually qualifies as an agent?
02 / RESEARCH AREA
Agent Architectures
Loops, graphs, supervisors and state machines
03 / RESEARCH AREA
Tools & Protocols
Interfaces are not permissions
04 / RESEARCH AREA
Modern Multi-Agent Systems
Coordination without assumed trust
05 / RESEARCH AREA
Memory & Context
Remembered data must keep its provenance
06 / RESEARCH AREA
Planning & Reasoning
A plan is a proposal, not proof
07 / RESEARCH AREA
Evaluation & Benchmarks
Reproducibility before leaderboard position
08 / RESEARCH AREA
Security & Containment
Minimize what an agent can touch
09 / RESEARCH AREA
Human Control
Humans need meaningful stopping power
10 / RESEARCH AREA
Agent Field Reports
Observable deployments, not miracle demos
FIELD REPORT / EDITION 001
Delegation is not permission.
A source-led examination of tool authority, model-controlled actions, operational risks and the approval boundaries that keep agents accountable.
Read the research note ↗BOUNDARY MODEL / CONCEPTUALNOT A BENCHMARK
01 Model proposes an actionUntrusted output / task context
02 Gateway checks scopePolicy · credentials · rate limits
03 Human approves protected writesRecorded decision · specific target
EDITORIAL CONTRACTRigorous by
Rigorous by
construction.
No vendor endorsements disguised as findings. No invented evaluations. No quietly changed conclusions.
Test against reality.
Use reproducible evaluations, baseline comparisons and independent outcome verification where technically and ethically feasible.
Preserve dissent.
Conflicting evidence and reported failures are reviewed. Material corrections are identified in a public edition history.
No live production agent data or benchmark results are represented on this launch site. The network visualization is conceptual; field reports will be distinguished from methodological commentary.