AI field guide / review
Review the language.
Recall the definition first, then open the term to check your answer. Progress is kept in your browser by the interactive page.
50 essential terms.
- Agent Concept · Agents & architecture
A system that selects actions, observes their results, and continues toward an objective.
- Harness Concept · Agents & architecture
The surrounding execution system: model calls, tools, state, context, permissions, and stopping rules.
- Harness engineering Concept · Agents & architecture
Improving the environment and controls around an agent, rather than only its prompt or model.
- Bounded autonomy Concept · Agents & architecture
Independent action within explicit permissions, budgets, scope, and stop conditions.
- Specification Concept · Goals, specs & plans
A precise description of required behavior, interfaces, or properties.
- Plan Concept · Goals, specs & plans
A proposed strategy and sequence for executing the work.
- Spec-driven development Pattern · Goals, specs & plans
A workflow in which explicit specifications guide design, implementation, and verification.
- Agent loop Pattern · Loops, critics & adversaries
Repeatedly assemble context, request an action, execute tools, observe results, and continue or stop.
- Replanning Concept · Loops, critics & adversaries
Revise the execution plan after new information, failures, or changed constraints.
- Adversarial agent Concept · Loops, critics & adversaries
An agent assigned to oppose, challenge, or attack a target within a defined setting.
- Adversarial loop Pattern · Loops, critics & adversaries
An informal pattern of challenge, revision, and rechecking; not one standardized algorithm.
- Critic Concept · Loops, critics & adversaries
A component that assesses a candidate and identifies weaknesses or potential improvements.
- Verifier Concept · Loops, critics & adversaries
A component that checks claims or properties against evidence and specified conditions.
- Repair loop Pattern · Loops, critics & adversaries
Modify a candidate or strategy in response to diagnosed failure, then try again.
- Stop condition Concept · Loops, critics & adversaries
An explicit rule for completion, failure, budget exhaustion, cancellation, or escalation.
- Subagent Concept · Multi-agent coordination
A delegated agent with its own task scope and often separate context or tool access.
- Handoff Pattern · Multi-agent coordination
Transfer responsibility or conversational control to another agent.
- Context engineering Concept · Context & memory
Select, structure, and update the information available to the model throughout execution.
- Compaction Concept · Context & memory
Replace detailed history with a smaller representation that preserves task-relevant information.
- Progressive disclosure Concept · Context & memory
Expose concise metadata first and fuller instructions or resources only when relevant.
- Tool calling Concept · Tools & output contracts
The model proposes a named operation and arguments; the runtime validates and executes the operation.
- Structured outputs Concept · Tools & output contracts
Model responses constrained to a supported output schema.
- Tool discovery Concept · Tools & output contracts
Find relevant tool capabilities at runtime instead of preloading every definition.
- Programmatic tool calling Concept · Tools & output contracts
Use model-generated code to orchestrate tool calls and process intermediate outputs.
- MCP Protocol · Protocols & interoperability
Model Context Protocol: a standard interface for exposing tools, resources, and prompts to AI applications.
- A2A Protocol · Protocols & interoperability
Agent2Agent: a protocol for communication between independently implemented agentic applications.
- ACP Protocol · Protocols & interoperability
Agent Client Protocol: standard communication between coding agents and editor or IDE clients.
- Skill Concept · Coding-agent internals
A reusable procedure or knowledge package that an agent can load when relevant.
- RAG Pattern · Retrieval & knowledge
Retrieval-augmented generation: fetch external evidence and use it to support a generated response.
- Embedding Concept · Retrieval & knowledge
A learned vector representation used for similarity and other downstream tasks.
- Hybrid search Concept · Retrieval & knowledge
Combine complementary retrieval methods, commonly lexical and dense vector search.
- Reranker Concept · Retrieval & knowledge
A component that re-scores an initial candidate set for relevance.
- Grounding Concept · Retrieval & knowledge
Tie generated claims or actions to relevant evidence and observations.
- Eval Concept · Evals & observability
A systematic test of a model or application's behavior against defined criteria.
- Trace Concept · Evals & observability
A structured record of model calls, tool actions, timing, errors, and related execution events.
- LLM-as-a-judge Concept · Evals & observability
Use a model to assess outputs or execution trajectories against a rubric.
- Task success rate Concept · Evals & observability
The fraction of evaluated tasks meeting the specified completion criteria.
- Cost per successful task Concept · Evals & observability
Total execution cost divided by successful outcomes, including the cost of unsuccessful attempts.
- TTFT Concept · Inference & performance
Time to first token: delay from a defined request start until the first generated token arrives.
- KV cache Concept · Inference & performance
Stored attention keys and values reused to avoid recomputing prior-token attention state.
- Prompt caching Concept · Inference & performance
A provider's mechanism for reusing eligible prompt processing; exact behavior and billing are platform-specific.
- Model routing Concept · Inference & performance
Choose a model based on task needs, constraints, policy, or observed difficulty.
- SFT Concept · Models, reasoning & training
Supervised fine-tuning: train on examples of desired inputs and outputs.
- LoRA Concept · Models, reasoning & training
Low-rank adaptation: learn low-rank parameter updates instead of updating all model weights.
- Inference-time compute Concept · Models, reasoning & training
Computation spent solving a request rather than training the model.
- Prompt injection Concept · Security & reliability
Untrusted content attempts to redirect a model away from the application's intended instructions.
- Least privilege Concept · Security & reliability
Grant only the access and capabilities necessary for a specific task.
- Idempotency Concept · Security & reliability
Repeated execution of the same operation has the same intended effect as one execution.
- Durable execution Concept · Security & reliability
Recover and continue a workflow across interruptions using persisted state or event history.
- Generative UI Concept · Generative UI & voice
Use model output to help determine which interface components and content a user sees.