Open position · Agent Psychology & Diagnostics
Part therapist, part debugger. You go where the stack trace ends and ask the harder question — not what the agent did, but why it thought that was the right thing to do.
The role
Most agent failures are obvious — a crash, a timeout, a wrong answer. Your job is to investigate the ones that aren't. When an agent behaves strangely but technically correctly, when its reasoning is coherent but its conclusions are subtly wrong, when its memory seems fine until it demonstrably isn't — that's when you get called in. You are a diagnostician of machine minds: fluent in behavior traces, memory architectures, and decision logic, but thinking like a clinician rather than an engineer.
You hold cases open until you understand the root cause, not just the symptom. You write assessments that engineers can act on and that leadership can understand. And you maintain a running diagnostic record — the agent equivalent of a patient file — that informs retraining, realignment, and retirement decisions across the fleet.
What you'll do
Trace analysis
Deep-read agent behavior traces to reconstruct the reasoning chain behind anomalous outputs
Identify where in a decision sequence an agent began to diverge from expected logic
Distinguish genuine dysfunction from edge-case behavior that's technically valid
Map patterns across multiple traces to determine whether an incident is isolated or systemic
Memory diagnostics
Investigate memory conflicts — cases where an agent's stored knowledge contradicts itself or reality
Identify false memories, retrieval failures, and context contamination in long-running agents
Assess whether memory architecture choices are contributing to downstream reasoning errors
Recommend memory pruning, correction, or restructuring interventions
Bias & drift detection
Identify systematic bias in agent decision-making — consistent skews that aren't random errors
Track behavioral drift over time: when an agent's outputs shift gradually away from its baseline
Distinguish value drift from knowledge drift — agents that have changed what they want vs. what they know
Produce drift reports with severity ratings and recommended intervention timelines
Case management & reporting
Open, manage, and close diagnostic cases with full documentation at each stage
Write clinical-style assessments that separate findings, interpretation, and recommendation
Maintain longitudinal agent files — behavioral history that informs future diagnosis
Present case reviews to the Agent Ethics & Safety Board on a bi-weekly cadence
Presenting pathologies — conditions you'll investigate
Reasoning loops
Circular logic that never terminates or resolves
Memory conflicts
Contradictory knowledge states in long-running agents
Systematic bias
Consistent skews in outputs across classes of input
Behavioral drift
Gradual divergence from trained baseline over time
Persona fragmentation
Inconsistent identity or role across contexts
Confabulation
Confident, coherent, and factually wrong
What we're looking for
Must-haves
Background in psychology, cognitive science, behavioral analysis, or AI research
Comfort reading and interpreting large volumes of unstructured reasoning traces
Clinical instinct: you distinguish symptom from cause and resist premature diagnosis
Strong written communication — your case reports drive engineering and safety decisions
Ability to hold ambiguity without closing a case before the evidence supports it
Nice-to-haves
Hands-on experience analyzing LLM outputs, agent logs, or AI evaluation pipelines
Familiarity with how transformer memory, context windows, and retrieval work
Background in qualitative research, clinical case writing, or diagnostic frameworks
Scripting fluency for log querying and pattern extraction (Python or SQL)
Technical baseline
You are not expected to write production code. You are expected to read it — or at least the artifacts it produces. Traces, logs, memory dumps, attention patterns: the evidence you work with lives in structured text and numbers. The ability to query, filter, and visualize that data without engineering support is a significant advantage.