Glossary — Agent Engineering Lab

Glossary

Plain, opinionated definitions of the terms the Lab keeps using.

Loop engineering: The 2026 practice of building the system that operates an agent instead of prompting it turn by turn: schedules and automations, parallel runs, maker-checker sub-agents, durable state. The orchestration is genuinely new. The loop it drives, and that loop’s stop problem, are not. See more →
Agent loop: The inner cycle a single agent runs: reason, act, observe, decide whether to continue, retry, stop, or escalate. Formalized as a control loop since the 1990s; ReAct is the modern language-model version. See more →
Harness: The runtime around a model: what enters the context window, how tools are dispatched, which hooks fire, how sub-agents, memory, and permissions work. The unit of evaluation is harness plus model plus task, not the model alone. See more →
Context engineering: Managing everything the model sees on a given call: what to include, compact, retrieve on demand, or drop. The discipline that replaced prompt engineering once context windows got large.
Maker-checker (verifier): A second agent that reviews the first agent’s work and acts as the loop’s stop condition. The evidence shows a model is a weak judge of its own kind of output, so the checker has to be external to the maker, a code executor, a test suite it did not write, or a model with a different objective. See more →
Stop condition: How a loop decides it is finished. The oldest unsolved problem in agent loops, named the horns of a dilemma in 1995. A stop condition the model self-grades is not a brake. See more →
Capability gate: An enforced limit on what an agent may do, checked outside the model: a destructive action that cannot run without an out-of-band approval the agent does not hold. The fix for a credential that was wider than the task.
Blast radius: How much damage one agent action can do. The question an incident review asks first: when this loop acts against something that matters, what is it allowed to touch, and who can prove afterward that it was allowed.
No-progress detection: Halting a loop when it repeats the same call without new information. A cheap, reliable fence against a stuck agent.
Iteration cap: A hard ceiling on loop turns so a runaway cannot burn unbounded budget. A safety net, not a primary stop condition.
Failure buckets: Categorizing evaluation failures by mode instead of collapsing them into one accuracy number. Two agents at the same aggregate score can fail in completely different, production-relevant ways. See more →
Workflow versus agent: Most agents are workflows in costume. Reach for an agent only when the task needs judgment a workflow cannot encode; the win condition is dynamic decision-making, not raw capability. See more →
Multi-agent versus router: Whether to split work across agents is a workload question, not an architecture preference. If the work decomposes into branches that do not see each other, multi-agent can help; otherwise a single agent or a router is cheaper and more reliable. See more →
Trace: The recorded sequence of what an agent did and why. If you cannot inspect the trace, you cannot debug, audit, or govern the agent. The trace is the truth.
Human-in-the-loop: Approval gates, escalation paths, and audit trails treated as first-class architecture, designed up front, not bolted on after legal complains.
Lethal trifecta: The security shape behind most agent incidents: access to private data, exposure to untrusted input, and the ability to act externally. Hold all three and a prompt injection becomes an exfiltration or a destructive call. See more →
Delegated authority: Authority an agent carries as a verifiable token that can only narrow as it passes down a chain of delegations, enforced outside the model. The missing layer under autonomous loops. See more →