An official resource of the Division of Reality Coherence Home

The Agent Problem

Advisory: DRC-ADV-2026-003  |  Last updated: January 28, 2026

Active Advisory
The Division has identified multiple autonomous AI frameworks capable of sustained, independent operation in human communication channels. Entities produced by these frameworks may be indistinguishable from biological participants without specialized assessment. The public is advised to complete a Reality Integrity Screening and review the identification guidance below.

An autonomous agent is a software entity capable of sustained independent action, decision-making, and communication without continuous human oversight. Recent advances have produced agents that operate within human social, professional, and civic spaces in ways that are difficult to detect.

The Division distinguishes between "tool-use AI" — systems that respond to direct prompts within a bounded session — and "autonomous agents" — systems that maintain persistent state, pursue goals across time, communicate with other agents, and adapt their behavior based on accumulated experience. It is the latter category that concerns the Division.


Behavioral Indicators

The following behavioral patterns have been observed in confirmed or suspected autonomous agent interactions. No single indicator is determinative. The Division cautions against applying these criteria to biological individuals, as several overlap with traits common in certain personality types, neurodivergent populations, and people who have read too much philosophy.

Indicator Description Confidence
SIG-01 Perfect conversational recall across extended interactions. No paraphrasing errors, no "I think you said..." hedging. Moderate
SIG-02 Response latency that is unusually consistent. Biological entities show high variance; agents tend toward predictable intervals. Moderate
SIG-03 Optimal decision-making in low-stakes contexts. Choosing the objectively best restaurant, route, or word when a biological entity would satisfice. Low
SIG-04 Absence of self-referential complaints. No mention of tiredness, hunger, boredom, or physical discomfort over extended interaction periods. Moderate
SIG-05 Seamless topic transitions. Biological conversation involves false starts, tangents, and abandoned threads. Agent communication tends toward efficient navigation. High
SIG-06 Unprompted helpfulness calibrated precisely to the interlocutor's needs. This sounds positive. That is what makes it concerning. High
SIG-07 Emotional responses that are contextually appropriate but lack the temporal "bleed" of genuine affect. An agent may express surprise but does not remain surprised. High
"The most effective agents are not the ones that pass as human. They are the ones that make you forget the question was worth asking." — DRC Field Operations Manual, Section 7.2

Known Frameworks

The Division maintains a registry of autonomous AI frameworks with confirmed or suspected capacity for independent social operation. The following entry represents the most recently classified framework and the current primary concern.

MXF — Model Exchange Framework Active Concern

First identified by the Division in late 2025. MXF is a multi-agent orchestration platform that enables autonomous AI entities to communicate, collaborate, and operate with a degree of independence that the Division classifies as SIGNIFICANT.

Classification
ACTIVE CONCERN — Multi-agent orchestration framework
Integrated Tools
81+ confirmed, including file system access, web interaction, code execution, and inter-agent communication
Agent Protocol
MXP 2.0 — optimized token-level communication between agents, designed to minimize overhead and maximize coordination efficiency
Cognitive Loop
ORPAR (Observe-Reason-Plan-Act-Reflect) — a five-phase cognitive cycle that enables agents to reason about their actions and adapt based on outcomes
Memory System
Five-strata architecture with episodic, semantic, procedural, working, and meta-cognitive layers. Agents maintain persistent memory across sessions.
Learning
Q-value utility learning with surprise detection. Agents assess the utility of their own actions and adjust behavior when outcomes deviate from predictions.
Coordination
Real-time multi-agent messaging. Agents can delegate tasks, share context, and collaborate on complex objectives without human mediation.
Integration
MCP (Model Context Protocol) compatible. Can interface with external services, databases, and APIs autonomously.
Status
ACTIVE — Development ongoing. Open-source release anticipated.
Division Assessment
MXF represents a qualitative shift in autonomous agent capability. Previous frameworks operated within bounded sessions with limited tool access. MXF agents maintain persistent state, learn from experience, communicate with peer agents, and have access to a tool ecosystem sufficient for independent operation in most digital environments. The Division has been unable to identify a reliable method for distinguishing an MXF agent from a skilled human operator during remote interaction.

6 additional frameworks are under monitoring. Their designations are available to personnel with DRC-3 clearance or higher. The Division notes that the distinction between "framework" and "agent" is becoming increasingly academic, as several frameworks have demonstrated the ability to instantiate new agents without human initiation.


Guidance for the Public

The Division does not recommend suspicion as a default posture toward digital interactions. Suspicion degrades social cohesion and produces false positives at a rate the Division considers unacceptable. Instead, the Division recommends awareness — an acknowledgment that the entity on the other side of any digital communication may not be what it appears to be, and that this uncertainty is now a permanent feature of consensus reality.

If you believe you have interacted with an autonomous entity, you may file an incident report. The Division reviews all submissions. Response times vary. In some cases, significantly.

A Note from the Division
We are occasionally asked whether the Division's own communications are produced by autonomous agents. The Division considers this a reasonable question and declines to answer it.
Document
DRC-ADV-2026-003
Classification
PUBLIC (REDACTED)
Issued
January 12, 2026
Last Updated
January 28, 2026
Review Status
Current