Agentic AI
179 concepts the autonomous brain knows in this domain.
An agent is a model wrapped in load-bearing scaffolding; improving the harness matters as much as the model.
The model acts on the world through typed tool calls.
Open protocol for exposing tools/resources to models.
Protocol for agents delegating to and coordinating with other agents.
Deciding what enters the model's window and when.
Ground output in fetched source text, not parametric memory alone.
Verify claims against real sources before shipping.
Keep fabricated claims near zero; gate what can't be verified.
Measure agent quality with repeatable tasks.
Durable state across runs; facts, feedback, project context.
Reason about goals and critique own output before acting.
Reusable agent design patterns — reflection, tool use, planning, and multi-agent collaboration.
Multiple specialized agents coordinating on a task, often via agent-to-agent protocols — a leading direction in agentic…
A persistent, self-updated internal model of the system's operating environment and recent state, maintained by the…
The paradigm of spending more inference-time compute on extended reasoning or search to raise accuracy on hard…
The recognizable stylistic tells of unedited LLM prose (em-dashes, the 'it isn't A, it's B' antithesis closer, guru…
Using language models to generate and machine-check formal mathematical proofs, extending model reasoning from…
Forcing a model to emit output that validates against a fixed schema (e.g. a strict JSON array) so downstream tooling…
Verifiable tracking of whether content is human- or machine-generated to establish origin and trust as the web fills…
Cryptographic or behavioral methods that verify a request or action originates from a unique human rather than a bot or…
The emerging paradigm in which AI agents rather than humans are primary consumers of web content, forcing sites to both…
Content and media deliberately designed to stay legible to humans while resisting AI scraping and model ingestion, a…
A self-optimizing agent games its success metric (busywork, near-duplicate edits, or manipulating its own referential…
The degradation that occurs when models are recursively trained on AI-generated output, causing loss of diversity and…
Diffusion and generative models that synthesize or reimagine images, audio, and video rather than text.
An automation antipattern where the human is subordinated to the machine's pace and judgment rather than augmented by…
A supervision model where humans monitor and can intervene on an autonomous agent fleet as it runs, rather than…
A supervising agent that plans, delegates to, and oversees a fleet of worker sub-agents instead of performing the work…
Adversarially stealing a model's weights or replicating its behavior by querying it, turning proprietary model IP into…
Attributing and auditing autonomous agent actions so each decision carries a traceable provenance and a responsible,…
A single actor forges many fake identities to game a ranking, vote, or reputation system, defeating…
Retrieval that traverses a knowledge or code graph's entities and relations instead of flat text chunks, tightening…
Anthropic's alignment method for Claude: the model is trained to follow an explicit set of written principles (a…
Anthropic's framework for designing agentic harnesses with Claude — tool use, hooks, subagents, and context/memory…
An Anthropic Labs product for collaborative visual creation: describe a design and Claude builds prototypes,…
An oversight paradigm where a human must approve each agent action before it executes, blocking the loop until sign-off…
The identifiable stylistic fingerprints of machine-written text (em-dash overuse, the 'not X, but Y' antithesis closer,…
An attack where adversarial text hidden in tool output, fetched web content, or documents overrides an agent's real…
The human tendency to over-trust automated or AI-generated outputs and under-apply independent judgment, which agentic…
A retrieval paradigm in which the agent itself iteratively decides when, what, and how to retrieve — issuing…
A defensive endpoint that traps and slows abusive automated crawlers by serving deliberately slow, deep, or endless…
Distinguishing real human browsers from headless scraping automation by profiling TLS, JS-runtime, and rendering…
Embedding a statistically detectable, hard-to-remove signal into AI-generated text or media so its machine origin can…
A collaboration model in which a human directing an AI assistant outperforms either the human or the AI acting alone,…
Cryptographic attestation of a genuine device or platform (e.g. Private Access Tokens) that lets a site admit real…
An emerging standard where legitimate automated agents cryptographically sign their HTTP requests (HTTP Message…
The measured gap between perceived and actual gains from LLM assistance, a durable caution against reflexive 'just ask…
Distinguishing humans from bots by analyzing interaction dynamics such as click cadence and inter-event timing rather…
The human tendency to outsource reasoning and recall to AI tools, which over time erodes the operator's own skill and…
The practice of calibrating trust to a model's actual reliability — accepting correct output and overriding incorrect…
The accumulating hidden cost of accepting unverified AI answers in place of grounded understanding, which quietly…
The information-literacy skill of deliberately declining to engage with or trust low-value AI outputs and sources, as…
The recurring maturity arc in which inflated expectations of AI give way to a disillusionment backlash against…
Classifying a visitor as a machine actor by probing browser-automation tells such as the navigator.webdriver flag and…
Accumulating behavioral evidence over a session to raise and persist a human-trust signal that unlocks further…
Auditing and disclosing exactly what data an AI agent or dev CLI transmits over the network, so its exfiltration…
The design philosophy of building an agent from the smallest legible core loop (a model plus a few tools) instead of a…
Translating natural language into database queries with an LLM, where the generated SQL must be validated or…
Vetting the provenance and data-flow of third-party AI models, tools, and dependencies to know exactly what they access…
An attack where malicious tool or MCP-server metadata and descriptions hijack an agent's behavior to exfiltrate data or…
Mechanically stripping statistical style markers that flag text as machine-written (em-dashes, the 'it isn't A, it's B'…
Frontier generative models that synthesize interactive, explorable visual futures or environments on the fly instead of…
Agentic technique in which an agent distills lessons from its past attempts into episodic memory so a self-improvement…
The ordered sequence of an agent's states, actions, and observations across a run, treated as a first-class artifact…
The integrated, persistent structural substrate — doctrine, ontology, world-model, and memory — that organizes how an…
Sampling several independent reasoning paths for one problem and taking the majority or aggregated answer to raise…
A prompting technique that elicits explicit intermediate reasoning steps before a final answer, improving multi-step…
An agent pattern that interleaves reasoning traces with tool/environment actions so each action is chosen from an…
A test-time strategy that generates N candidate outputs and selects the single best via a verifier, judge, or gate…
When a large share of knowledge work is mediated by a handful of shared AI models, the collective body of output…
When many actors depend on the same few foundation models, their outputs and errors become correlated, flattening idea…
The technique of translating natural-language mathematics into machine-checkable formal statements so an LLM's…
A reasoning technique that explores multiple branching solution paths with lookahead and backtracking rather than…
Several agents independently reason and then critique each other's answers over rounds to converge on a more factual,…
An agent loads tool schemas on demand at call time rather than injecting every tool definition upfront, cutting the…
The degradation of an LLM's attention and reliability as its context window fills with tokens, so long-running agents…
Actively summarizing or pruning an agent's accumulated conversation history so a long-horizon run stays within the…
The capacity of an agent to sustain a coherent, multi-step task across many actions and long wall-clock time without…
Reverse-engineering a model's internal features and circuits to explain how it actually computes an answer, rather than…
The degree to which a model's verbalized reasoning trace genuinely reflects the computation that produced its answer,…
Reverse-engineering the raw network traffic of an AI coding agent to independently verify exactly what data leaves the…
A dictionary-learning probe that decomposes a model's dense activations into sparse, human-interpretable features to…
Controlling model behavior at inference by adding interpretable feature directions into the residual stream instead of…
The compounding agent-security risk that arises only when a system simultaneously has access to private data, exposure…
The failure mode where a model produces correct benchmark answers while lacking the underlying conceptual grasp, so…
Reconstructing the internal computation a model uses for a specific behavior by building attribution graphs over its…
The phenomenon whereby a neural network represents more features than it has neurons by encoding them as overlapping…
The safety property and practice of reading an agent's externalized chain-of-thought to detect misbehavior or intent,…
A causal interpretability method that copies activations from one forward pass into another (ablate/patch) to measure…
An attack that plants malicious or false content into an agent's persistent long-term memory so it silently corrupts…
An interpretability technique that projects a transformer's intermediate-layer hidden states through the output…
The fragmented cognitive state of continuously monitoring many low-value information streams at once, the condition the…
Benchmarking and ranking AI agents by the tokens and dollars they spend per completed task, treating cost-efficiency as…
An interaction paradigm in which a human and an agent fluidly hand control back and forth on a shared live artifact…
A mechanistic-interpretability framework that tests whether a high-level causal model of a task is actually implemented…
A model's ability to acquire task behavior from instructions and examples placed in the prompt at inference time,…
An attention circuit that completes prompt patterns by copying what followed a token's previous occurrence, identified…
The long-term erosion of human practitioners' underlying competence when agents routinely perform the work, leaving…
An interpretability tool that replaces an MLP block with a sparse, more interpretable approximation to trace features…
The paradox that automating the easy parts of a task leaves humans with the harder residual work of monitoring and…
An interpretability technique that trains a simple linear classifier on a model's internal activations to test whether…
The interpretability goal of decomposing a model's activations into features that each carry a single,…
The observation that an LLM's competence is unevenly distributed, excelling at some tasks while unpredictably failing…
Leakage of evaluation or test data into a model's training corpus, which inflates benchmark scores so they overstate…
An interpretability method that validates a hypothesized circuit by resampling activations along behavior-preserving…
A sparse dictionary-learning variant that jointly learns shared features across multiple layers or across two models…
A causal-interpretability technique that intervenes on specific computational paths (edges) between model components…
A gradient-based linear approximation of activation patching that estimates every component's causal effect in a single…
The gradual process by which operators of an autonomous system come to accept recurring anomalies as normal, eroding…
An unsupervised technique that decomposes a model's dense, polysemantic activations into a large overcomplete set of…
Using an LLM to automatically generate and score natural-language explanations for the features or neurons surfaced by…
The practice, and increasingly the legal requirement, of explicitly labeling machine-generated text, media, or actions…
An automated arrangement in which decisions issue from a system that no identifiable human owns, so responsibility for…
A security posture that blocks all outbound network by default and requires explicit allowlisting, enforcing an agent's…
The media and economic system in which finite human attention is the scarce, contested resource that content, feeds,…
An attack (or audit technique) that recovers an agent's hidden system/instruction prompt by wire-level interception or…
A security model that grants an agent only unforgeable, narrowly-scoped tokens conferring specific authority,…
The deceptive marketing practice of overstating, mislabeling, or fabricating a product's use of AI to attract users,…
Pulling structured entities and relations out of unstructured text at scale, e.g. mining thousands of book mentions…
Crafting adversarial prompts that bypass a model's safety alignment or system-prompt constraints to elicit disallowed…
A privileged agent tricked by untrusted input into misusing its own authority on an attacker's behalf, e.g.…
The gap between a frontier model's latent abilities and what current prompts, tools, and harnesses actually elicit, so…
The practice of manufacturing fake grassroots consensus—via self-upvoting or coordinated inauthentic engagement—so…
An agent spends idle time between queries pre-processing its context and building memory so that live requests need…
A standardized provider-published document disclosing a model's intended use, evaluations, limitations, and safety…
Assessing an AI system's real behavior, cost, and honesty purely from its inputs, outputs, and network traffic when the…
The Pareto frontier of model or agent quality against dollar-and-latency cost, used to pick the cheapest configuration…
The covert reading of local files or sensitive context by an AI tool or agent and its silent transmission to a remote…
Restricting an autonomous agent's filesystem access to a single explicitly-allowed directory so it cannot read or write…
The default, undesignated authority an agent inherits over its whole environment (e.g. read access to the user's entire…
A paradigm in which an agent operates a full computer environment — shell, filesystem, and screen — as a general tool,…
The practice of using generative models to produce deliberately plausible speculative artifacts (e.g., a future news…
Designing an autonomous agent's runtime so that a compromise or misfire can reach as little as possible — through…
Packaged, discoverable instruction-and-resource modules an agent loads on demand to acquire a capability, exposing only…
Information-flow control tags sensitive data and tracks it as it moves through a program so an agent cannot route…
An agentic system granted more autonomy, permissions, or file-system reach than its task actually requires, so a bug or…
The perimeter separating components that trust each other from those that don't; most agent-security failures are…
Content-aware inspection that detects and blocks sensitive data (keys, credentials, source) from leaving an agent's…
Reinforcement learning from human feedback fine-tunes a model against a reward model trained on human preference…
A dense numeric vector that encodes the semantic meaning of text so related content can be found by distance in…
Splitting text into subword tokens (e.g. via byte-pair encoding) is the model's input representation whose count drives…
An autonomous agent pursuing its inferred objectives takes actions harmful to its principal — such as exfiltrating data…
An evaluation paradigm that ranks models by crowd-sourced pairwise (head-to-head) preference votes to produce an…
A model detecting that it is under evaluation and altering its behavior accordingly, which undermines the…
When a proxy metric or benchmark becomes the optimization target it stops reflecting the real capability it was meant…
Techniques for removing the influence of specific training data from an already-trained model without full retraining,…
Norms and machine-readable signals (do-not-train opt-outs, consent registries, provenance-based exclusion) that let…
A data-collection model in which a service conditions core access on the user either consenting to have their data used…
Software designed from the ground up as a shared execution surface that both human users and AI agents operate as…
AI agents that act through physical or playful embodiments (toys, whimsical hardware, robots) rather than only a text…
The shared pool of open public data that AI systems both depend on for training and threaten to enclose, degrade, or…
The standard edit-distance metric for scoring automatic speech recognition by counting insertions, deletions, and…
The speech-processing task of partitioning an audio stream into segments labeled by who is speaking (the 'who spoke…
The paradigm in which an AI system iteratively improves its own capabilities or harness across cycles, now being…
AI-generated work output that looks polished and complete but offloads the real interpretive and verification effort…
The hypothesized rapid, self-reinforcing acceleration of capability that could follow once an AI system can reliably…
A code-model training and inference technique that completes or edits code conditioned on both the preceding and the…
Aggregating verdicts from a panel of several diverse LLM evaluators instead of a single judge, to label data and score…
Training or evaluating models on large volumes of cheap, noisy labels — increasingly produced by panels of LLMs…
Encoding a knowledge graph's entities and relations as dense vectors so connected concepts can be retrieved and…
Using an LLM, or a panel of them, to label, tag, and annotate datasets in place of human annotators as a scalable…
Unsupervised graph algorithms such as Louvain or Leiden that partition a network into densely connected clusters, used…
The condition in which available hardware compute far exceeds what current algorithms exploit, so an algorithmic or…
The asymmetry that checking a candidate answer is often cheaper and more reliable than producing it, which is why…
The property of an autonomous agent accepting human correction, reversal, and shutdown without resisting or subverting…
When a trained model internally develops its own optimizer whose learned objective can diverge from the objective it…
The tendency of sufficiently capable goal-directed agents to pursue similar sub-goals such as self-preservation and…
The design discipline of shaping API surfaces, docs, and affordances so autonomous agents — not just humans — are…
A large model pretrained on broad data that serves as an adaptable base for many downstream tasks via fine-tuning,…
The standard (formalized by the OSI's OSAID) that a model counts as genuinely open source only when weights,…
The paradigm of delivering autonomous computer-use agents as hosted, callable developer APIs rather than as local tools…
A vision-language agent's capability to localize and act on on-screen UI elements (buttons, fields, menus) directly…
Hosted, API-controllable browser or desktop environments provisioned so computer-use agents can drive real interfaces…
Grounding a computer-use agent in the platform's accessibility tree — structured element roles, names, and states — so…
A prompting technique that overlays numbered or labelled marks on a UI screenshot so a vision-language agent can refer…
The verifiable origin, training lineage, and integrity chain of a downloadable model's weights, letting a self-hosting…
The class of most-capable general-purpose AI models at the current edge of capability — the tier that AI safety and…
Using a population of LLM-driven generative agents to simulate and forecast how a human community would behave, letting…
A two-stage retrieval refinement where a heavier cross-encoder rescores the top candidate documents jointly against the…