AI field guide / distinctions
Similar words. Different jobs.
Short distinctions for terms that often get used as if they meant the same thing.
Agent Harness
The agent selects actions. The harness makes those actions executable and controlled.
A system that selects actions, observes their results, and continues toward an objective.
Skill Tool calling
A skill supplies a procedure. A tool call requests an operation.
A reusable procedure or knowledge package that an agent can load when relevant.
Specification Plan
A specification defines required behavior. A plan describes how to get there.
A precise description of required behavior, interfaces, or properties.
Critic Verifier
Critique identifies potential flaws. Verification checks defined claims against evidence.
A component that assesses a candidate and identifies weaknesses or potential improvements.
A component that checks claims or properties against evidence and specified conditions.
Adversarial agent Adversary
A reviewer can challenge work cooperatively; an adversary has an opposing objective.
An agent assigned to oppose, challenge, or attack a target within a defined setting.
An actor or process whose objective conflicts with the target system's objective.
Retry loop Repair loop
Retry repeats an operation. Repair changes the candidate or strategy.
Repeat an operation after failure, often without changing the underlying candidate or strategy.
Modify a candidate or strategy in response to diagnosed failure, then try again.
MCP A2A
MCP exposes tools and context. A2A connects independently implemented agentic applications.
Agent2Agent: a protocol for communication between independently implemented agentic applications.
ACP AG-UI
ACP connects editors and coding agents. AG-UI carries agent-to-interface events.
Agent Client Protocol: standard communication between coding agents and editor or IDE clients.
Prompt engineering Context engineering
Prompting is instruction design. Context engineering manages the full information state.
Design instructions, examples, and output requests to guide model behavior.
Select, structure, and update the information available to the model throughout execution.
JSON mode Structured outputs
Valid JSON is not the same as conformance to your declared schema.
An API mode targeting valid JSON, without necessarily enforcing your complete business schema.
Prompt caching Long-term memory
Reusing input computation is not remembering facts across tasks.
Information retained and retrieved across tasks or conversations.
TTFT End-to-end latency
A fast first token is not the same as fast task completion.
Time to first token: delay from a defined request start until the first generated token arrives.
Total request or task duration, including queues, model calls, tools, retrieval, and retries.
Git worktree Sandbox
A separate working directory is not an enforced security boundary.
A separate working directory linked to a repository; it is not a security isolation boundary.
An execution environment with enforced isolation and access or resource restrictions.
Faithfulness Correctness
An answer can accurately repeat evidence that is itself wrong.
Whether an answer remains supported by the evidence supplied to it.
Whether an answer or action satisfies the actual problem's requirements or truth conditions.
Durable execution Idempotency
Recovering a workflow does not by itself prevent duplicate external effects.
Recover and continue a workflow across interruptions using persisted state or event history.
Repeated execution of the same operation has the same intended effect as one execution.
Mixture of experts Agent team
Internal neural-network experts are not independently orchestrated agents.
A model architecture routing computation through a subset of expert components.
A collection of agents with assigned roles and a defined way to coordinate.