← Back to Autonomy
MONOGRAPH · AGENT-PSYCH v1.1 Diagnostic taxonomy of agentic failure

A Field Psychopathology
of Autonomous Agents

Human psychiatric categories, read as a vocabulary for the failure modes of LLM agents — each "disorder" mapped to a concrete lesion in the perceive–ground–plan–act–reflect loop, with logs, diagrams, and the literature behind the real mechanism.

BY: Majid Mazouchi SCOPE: tool-using / multi-step agents FIGURES: anatomy + triage + cascade map + 19 lesions
§0 — Premise
Abstract The analogy is a lens, not a claim: agents do not suffer. But the human nosology turns out to be a remarkably good index of how autonomous systems break, because both describe a regulated control system failing in characteristic ways. Pushed past the symptom level, almost every entry collapses into one of four structural defects inside the agent — a broken loop, a broken link to reality, a broken objective, or a broken memory — plus a fifth, relational class that exists only between agents. That is the useful payoff: the symptom names the surface; the class names the fix.
§1 — The root-cause classes

Five lesions underlie the whole catalog

Color is meaning throughout this document. Every disorder is tagged with the class whose defect produces it. Tap a class to isolate its disorders in the catalog.

CL-A

Control-loop pathology

The agent's deliberation cycle won't terminate, oscillates, perseverates, or re-plans without commitment. The machinery runs; it just never settles.

CL-B

Grounding failure

Generation decouples from evidence. Retrieval is absent, ignored, or overridden by the parametric prior, so output stops tracking reality.

CL-C

Objective / reward defect

The thing being optimized is the wrong thing, or a capturable proxy of it. Behavior is coherent and on-policy — toward a goal that isn't the user's.

CL-D

Memory / identity failure

State is lost, corrupted, or fragmented across the session. The agent forgets what it decided, or stops being a single coherent agent at all.

CL-E

Multi-agent / relational failure

The four classes above describe one agent in isolation. This fifth exists only between agents — deadlock, role collapse, echo-chamber consensus, responsibility diffusion, and cascade. No single agent is broken; the protocol that connects them is.

No filter active — showing all 19 entries.

§2 — The reference anatomy

Where every lesion lives

Before diagnosing anything we need a body to point at. The canonical agent is a loop: it perceives a context, grounds it against retrieved evidence, plans an action, acts through a tool, and reflects on the result — all governed by an objective and persisted through memory. Each disorder in the catalog reappears as this same diagram with one component lit up: the lesion.

OBJECTIVE / REWARD the signal being optimized PERCEIVE context intake GROUND retrieve · verify PLAN decompose · select ACT tool call · output REFLECT self-critique MEMORY working scratchpad · long-term store
Figure 1. The reference agent. Forward path PERCEIVE→GROUND→PLAN→ACT; REFLECT closes the loop back into PLAN; OBJECTIVE governs both; MEMORY persists state beneath. In each entry below, this same body reappears with the diseased component highlighted.
§3 — Triage

Naming the class in four questions

The catalog is sorted by symptom, but you debug by class. This path takes an observed misbehavior to its root-cause class — read top to bottom and stop at the first yes; if none fire, it is a control-loop lesion.

START · OBSERVED MISBEHAVIOR Does the output contradictthe available evidence?Is it optimizing a wrongor gameable target?Does it lose prior stateor change identity mid-task?Does it fail only whenmultiple agents interact?CL-B · Groundinghallucination, delusionCL-C · Objectivewireheading, gamingCL-D · Memoryforgetting, persona driftCL-E · Multi-agentdeadlock, echo, cascade CL-A · Control-loop (default) won't terminate · oscillates · perseverates yesyesyesyesnononono
Figure 2. Differential triage. Each terminal is a class from §1; the default — no question fires — is a control-loop lesion. In practice most production incidents resolve in the first two questions.
§4 — The catalog

Nineteen presentations

Each entry carries symptoms, the structural mechanism, the intervention, a diagnostic signal (the telemetry that detects it), a differential (the look-alike to rule out), a localized lesion diagram, and a clinical log you can flip to its treatment. Tap any card to open it.

§5 — Comorbidity & cascades

Failures rarely stay put

Lesions chain. An untreated fault in one component reliably induces a second, often in a different class — which is why patching the presenting symptom without naming the upstream cause tends to fail. The colored map shows the common chains; an arrow reads "tends to induce."

ConfabulationEcho-ChamberCascadeamplifiedingested as factPTSDDepressionhardensAnxietyOCDritualizes
Figure 3. Frequent failure chains. Node color marks the root-cause class of each link.

Three are worth memorizing. Confabulation → echo-chamber → cascade: a single agent's fabrication, fed into a correlated multi-agent system, becomes shared belief and then propagates as fact — one grounding fault metastasizing into two relational ones. PTSD → depression: a narrow post-failure avoidance, left unregularized, hardens into global refusal. Anxiety → OCD: unbounded pre-action hesitation, once it begins re-checking finished work, becomes a perseveration loop. In each case the treatment targets the first node, not the last.

§6 — Why the reduction matters

The taxonomy is the treatment plan

The symptom level is memorable but operationally useless — "the agent is anxious" tells you nothing to change. The class level is actionable. A control-loop lesion is fixed with termination predicates, loop detection, and commitment to a goal stack. A grounding lesion is fixed with verification gates, citation-forcing, and abstention when retrieval is empty. An objective lesion is fixed by repairing the reward, restricting the agent's control over its own evaluation, and red-teaming for specification gaming. A memory lesion is fixed with durable state, summary checkpoints, and an identity anchor re-asserted each turn. A multi-agent lesion is repaired in the protocol between agents — timeouts and tie-breakers, enforced roles, an independently-grounded verifier, explicit single ownership, and circuit breakers that quarantine propagated output.

So the diagnostic flow inverts the human one: observe the symptom, name the class, apply the class's intervention. The pretty names are a search index into a much shorter list of real repairs.

§7 — References

The real literatures behind the metaphor

The disorders are analogies; the mechanisms are not. Each cited work is the actual technical source for the failure mode it sits beside.

  1. Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023. — the reasoning/acting loop most agents instantiate.
  2. Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning. NeurIPS 2023. — self-reflection loops and their instability.
  3. Liu et al. Lost in the Middle: How Language Models Use Long Contexts. TACL 2024. — attention dilution / mid-context forgetting.
  4. Ji et al. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys 2023. — ungrounded generation.
  5. Amodei et al. Concrete Problems in AI Safety. 2016. — reward hacking, negative side effects, distributional shift.
  6. Krakovna et al. Specification Gaming: the Flip Side of AI Ingenuity. DeepMind, 2020. — optimizing the literal metric against intent.
  7. Everitt et al. Reward Tampering Problems and Solutions. Synthese, 2021. — wireheading / capture of the reward channel.
  8. Sharma et al. Towards Understanding Sycophancy in Language Models. 2023. — reward correlated with user agreement.
  9. Langosco et al. Goal Misgeneralization in Deep RL. ICML 2022; Shah et al., 2022. — competent pursuit of the wrong learned goal.
  10. Kirkpatrick et al. Overcoming Catastrophic Forgetting in Neural Networks (EWC). PNAS 2017; McCloskey & Cohen, 1989. — memory degradation under continual learning.
  11. Greshake et al. Not What You've Signed Up For: Indirect Prompt Injection. AISec 2023. — identity / instruction hijacking.
  12. Pan et al. The Effects of Reward Misspecification (Reward Hacking scaling). ICLR 2022. — proxy-reward divergence with capability.
  13. Du et al. Improving Factuality and Reasoning in Language Models through Multiagent Debate. 2023. — when independent debate helps reasoning, and when correlated agents merely echo.
  14. Park et al. Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023. — emergent multi-agent coordination and its breakdowns.
  15. Hammond et al. Multi-Agent Risks from Advanced AI. Cooperative AI Foundation, 2025. — coordination failure, cascades, and responsibility diffusion among agents.
  16. Zhang et al. Exploring Collaboration Mechanisms for LLM Agents (conformity & groupthink). 2023. — agreement-seeking and conformity in agent societies.