Terms from the field.

The effort a person or organization must invest to use a new tool productively. The higher the barrier, the slower the adoption.

The degree to which AI-generated output conforms to the constraints of a design system. A potential DesignOps metric for quality assurance of agent work.

A persistent, evolving state that gives AI agents context across sessions. Not to be confused with the context window (volatile) or RAG (stateless).

A working model where AI agents are integrated into team processes not as passive tools but as autonomous partners with their own context, session history, and decision-making capacity.

Software systems that act autonomously, navigate via LLM steering, and execute tasks. Differ from human users (GUIs) and traditional computer programs (APIs) through human-like behavior at computer-like execution speed.

An AI agent is an artificial intelligence system that autonomously executes multi-step tasks — not just providing single answers, but planning, acting, evaluating results, and iterating. Unlike a chatbot, an agent can use tools, write and execute code, edit files, and interact with external systems.

Research field addressing moral questions of AI system deployment. Encompasses technical approaches (algorithmic fairness, explainability), regulatory frameworks (EU AI Act), and philosophical groundwork. Internally heterogeneous, no unified methodological approach.

A public database of documented AI security incidents, curated via Stanford HAI. Foundation for the incident curve in the visualization. Licensed under CC BY.

The process of ensuring AI systems act in accordance with human values, intentions, and safety requirements. Goal: the system reliably does what humans want — even in unforeseen situations.

Hypothetical design pattern: AI systems that work in the background (updating context, checking consistency, preparing suggestions) and only move to the center of attention when something relevant occurs. A counter-proposal to the dominant chat paradigm.

System that automatically generates a pool of prompt candidates, evaluates them via scoring function, and selects the best. Demonstrates that LLMs can develop prompts at or above human level.

A hypothetical AI that surpasses human intelligence across all domains — not just narrow tasks like chess or coding, but generally.

Design philosophy formulated in 1995 by Mark Weiser and John Seely Brown (Xerox PARC). Core idea: The best technologies require minimal attention and move fluidly between the periphery and center of perception.

Technique where the model thinks step by step before responding. Improves accuracy on complex tasks — math, logical reasoning, analysis.

A project-level instruction file read by Claude Code at session start. Functions as the “system prompt” for a repository — encoding identity, conventions, and navigation.

2023 Princeton framework defining four memory types for AI agents — Working, Procedural, Semantic, Episodic Memory — continuing the cognitive science tradition beginning with Atkinson/Shiffrin.

Framework from Princeton (2023) defining four memory types for AI agents — derived from the cognitive SOAR architecture. The field’s lingua franca.

Repeatable patterns for teamwork — here specifically for interaction between humans and AI agents. They address the layer between individual prompts and organizational governance.

Jürgen Habermas’s term (1981) for mutual understanding under conditions that enable equal participation for all involved: comprehensibility, the possibility of objection, no structural asymmetry.

A Markdown wiki that an LLM generates from raw data and incrementally maintains — with summaries, categories, and backlinks. Read-heavy, with the LLM as maintainer.

The computational capacity required to train and run AI models. Typically measured in GPU-hours. Concentration of compute among a few actors is a central geopolitical issue.

Design of the information environment in which AI agents operate — file structures, documentation formats, conventions, retrieval paths. In the DesignOps context: the infrastructure that enables agents to work with design systems.

Tools like Gitingest or DeepWiki that prepare information for efficient LLM and agent processing. Address the problem that existing web formats (HTML) are poorly suited for AI consumption.

A phenomenon where an AI agent’s internal state representation increasingly diverges from reality over long sessions — similar to the “telephone game” effect in human communication.

Context Engineering refers to the ability to provide an AI agent with the right context — all the information, rules, and constraints it needs to solve a task well. The better the context, the better the result. In agent-driven workflows, this competency increasingly replaces manual coding as the engineer’s central skill.

Continuous Deployment means that code changes are automatically delivered to the production environment without a manual release step — as soon as all automated tests pass. Combined with feature flags and observability, it enables fast iteration cycles with controllable risk.

Independent actors develop the same solution to the same problem — not through coordination, but through identical selection pressures. In biology: wings in birds and bats. Here: Markdown + Git + LLM across different practitioners.

Tradition of social philosophy, originating at the Institute for Social Research in Frankfurt in the 1920s. Combines social analysis with the normative goal of human emancipation. Works with immanent critique rather than external ideal standards.

A software process that runs in the background without direct user interaction. In the KAIROS context: an agent that autonomously executes tasks after a session ends.

The design of data visualizations and data-driven interfaces — from charts to dashboards to interactive exploration tools.

A mathematical function that reduces a memory’s relevance score over time — imitating biological forgetting. Semantic similarity × exponential time decay.

Optimization of processes, tools, and structures that design teams need to work effectively and at scale. Encompasses People, Process, and Craft/Tool Operations.

Horkheimer and Adorno’s core thesis (1944): reason and rational control can turn into their opposite — into domination, normalization, oppression — when they become ends in themselves rather than means for human liberation.

Access to users through existing channels and products. Google reaches billions via search, Meta via social media — OpenAI has no comparable channel and must acquire users directly.

The time interval at which measurable AI model capabilities double. According to METR data, currently at approximately 7 months, trending toward 4 months.

Framework that treats prompt pipelines as declarative programs and automatically optimizes them at compile time. Replaces manual prompt tuning with systematic optimization.

In the context of Critical Theory: the expansion of human agency and self-determination — as a benchmark for social systems. In the AI context: the question of whether a system expands judgment or replaces it through dependency.

A graphical representation of how intensively and frequently users engage with a product. A “flat” curve means: many users, but shallow usage depth — most use the product rarely and superficially.

Logs of past experiences: conversation logs, action sequences, few-shot examples. The agent’s “autobiographical memory.”

A feature flag is a toggle in the code that allows new features to be selectively enabled or disabled — without redeploying the code. Teams use feature flags to release new features to a small group of users first and instantly roll back if problems occur. In agent-driven workflows, they serve as a central safety net for continuous deployment.

A fractional engagement means that an expert or team doesn’t work full-time for a client but contributes a defined share of their capacity — for example, 2 days per week. Common in consulting, design, and strategy. As AI agents take over expert tasks, the pricing and scoping of such engagements change fundamentally.

An organizational model in which multiple Fractional Design Leaders form a network and share infrastructure — design systems, knowledge base, methods, AI tooling. Scales the model beyond the individual.

Strategic design guidance in the fractional model: design system governance, team building, DesignOps, product design strategy — on a part-time basis as an embedded leadership role.

A leader who works on a part-time basis for multiple companies simultaneously. Takes on strategic responsibility but is not permanently employed. Typical: 1–3 days per week, self-employed.

Method of Critical Theory: a system is measured not against external norms but against its own promises and claims. Targets the tension between self-representation and reality.

Internal codename for an unreleased Claude Code feature: a background daemon that continues working autonomously after a session ends, including nightly memory consolidation.

A prompt that does not directly solve a task but generates, evaluates, or optimizes other prompts. Shifts work one abstraction level up — from content to structure.

METR (Model Evaluation and Threat Research) is an organization that systematically measures how long real-world tasks take that AI models can solve end-to-end without human help. METR data shows an accelerating doubling rate of AI capabilities and is one of the few quantitative sources for tracking the performance progress of AI agents.

An independent research organization that evaluates AI models for capabilities and risks. Provides standardized benchmarks for task complexity and agent performance.

A strategic competitive advantage that durably protects a company from competition — like a castle moat. Controversial for AI models: if models are interchangeable, the moat is absent.

Organization method by Tiago Forte for personal knowledge management. Four categories based on actionability, not topic. Foundation of many knowledge management systems now being combined with LLM agents.

The two modes of attention in Weiser/Brown’s model. Periphery = unconscious perception (informs without cognitive cost). Center = conscious focus (requires active attention). Good technology shifts seamlessly between both.

A strategic approach where a company opens its technology as a platform for others to build on. Examples: App Store (Apple), GPT Store (OpenAI). Goal: value creation through an ecosystem rather than a single product.

The observation that the growing number of AI ethics guidelines does not lead to more ethical action. The existence of principles creates the impression of solved problems while the structural causes remain unchanged.

Codified behavioral rules: system prompts, tool definitions, decision logic. An agent’s “muscle memory.”

Hypothetical role profiles from a stronger generalization of today’s product roles (PM, PD, Dev). Thinker = strategy, framing, decision. Builder = implementation, prototyping, iteration.

Control process maintaining information in working memory. Maintenance rehearsal involves simple repetition; elaborative rehearsal links information to existing knowledge — only the latter reliably produces long-term storage.

The set of practices ensuring software artifacts don’t contain unintended content when deployed (source maps, debug symbols, internal references). In context: the weakest link in Anthropic’s security architecture.

A training method that uses human evaluations to guide an AI model’s behavior. Goal: the model should give helpful, honest, and harmless responses.

The Software Development Lifecycle describes the classic sequence of phases in software development: gather requirements, design the system, write code, test, review, deploy, and monitor. Traditionally, these steps run sequentially, with clear handoff points between each phase. In practice, many teams are seeing these phases dissolve — especially when AI agents execute large parts of the cycle simultaneously.

The ability of an AI system to contribute to its own improvement — such as debugging its own training or managing its own deployment. GPT-5.3 Codex is considered the first documented example.

Iterative process: the model generates a response, critiques it, and produces an improved version. Repeatable until quality threshold is met.

Accumulated factual knowledge: user preferences, extracted facts, knowledge bases. Grows over time.

Initial memory stage holding sensory impressions for fractions of a second. Unlimited capacity, extremely brief retention. Filters by relevance before information reaches working memory.

A working mode where knowledge and progress are bound to individual chat sessions. Without explicit persistence (in files, issues, commits), context is lost when a session ends.

A self-directed experiment outside the official product roadmap. Serves exploratory innovation in environments with high uncertainty about future feasibility.

Karpathy’s term for the third software paradigm: natural language as programming language. Prompts replace code (1.0) and trained weights (2.0) as the primary programming interface.

A debug file that maps compiled/minified JavaScript back to the original TypeScript source code. Not intended for production releases — its publication exposes the complete source code.

Sprint Planning is a ritual from agile software development (Scrum) where a team plans which tasks to complete in the next one to four weeks. Tasks are estimated, prioritized, and assigned. In agent-driven workflows, Sprint Planning loses relevance because the cycle from idea to delivery is drastically shortened.

Machine-readable project information in versioned files (Markdown, YAML, JSON) — as an alternative to ephemeral chat history. Serves as the shared communication medium between human and agent.

Test-Driven Development is a development method where automated tests are written first, followed by the code that should pass those tests. The cycle: write test, write code, pass test, clean up code. With AI agents, TDD increasingly becomes default behavior — the agent generates tests and code simultaneously.

Concept by Andrew Feenberg (1991). The implicit value decisions inscribed in a technical system’s architecture — before any explicit ethics policy. Makes visible that “neutral” design decisions have political and social consequences.

Method that treats prompt optimization like gradient descent — using natural language feedback instead of numeric scores. Published in Nature (2025).

1968 model proposing three serial memory stages: sensory register, short-term memory, long-term memory. Historically influential; psychologically outdated.

Author of “Building a Second Brain” and developer of the PARA method. One of the most influential thinkers in personal knowledge management — his framework is increasingly being adapted as a basis for AI-augmented knowledge systems.

A tight loop describes a very short, tightly coupled feedback cycle between action and feedback. In the context of agent-driven software development, it refers to the cycle: formulate intent, agent builds, observe result, adjust — without the long wait times of traditional development processes.

Term coined by Mark Weiser (1991) for pervasive, invisible computing technology. The theoretical framework from which Calm Technology emerged.

The current conversation context — what the agent is actively “thinking” about. Corresponds to the context window. Fast but volatile.