Glossary.
Plain definitions for the terms we use, written the way we use them, not the marketing version. We're adding entries as we go.
Shadow Deployment
Running a new AI agent version in parallel with the production agent on real incoming requests, with the new version's output discarded rather than served to users. Shadow mode surfaces systematic differences in cost, latency, or quality that unit tests miss. Typically runs 24+ hours before advancing to canary testing.
Canary Rollout
Gradual traffic routing to a new agent version, starting at 1-5% of real traffic with hard spend caps and automated error-rate gates. If error rate or cost per request stays within tolerance, traffic expands in stages; if not, the rollout reverts immediately.
Trajectory Quality
An evaluation of the intermediate steps in a multi-step agent workflow — tool selection, reasoning quality, error recovery, escalation decisions — not just the final output. A tool-calling error early in a trajectory silently corrupts every downstream step, which final-answer accuracy alone won't catch.
Agent Observability
Instrumentation that reconstructs an agent's full execution path — tool calls, memory reads and writes, reasoning steps, and handoffs between agents — rather than just input, output, latency, and cost the way standard LLM-call observability does.
Circuit Breaker (Tool Calls)
A safeguard that halts an agent's retry loop after a fixed number of identical tool-call failures, commonly three, instead of retrying indefinitely. Prevents cost runaway and stops a single malformed or hallucinated tool call from corrupting every downstream reasoning step.
A2A Protocol
Agent-to-Agent (A2A) is an open protocol, originated by Google and now under the Linux Foundation, for peer-to-peer communication between independent agents owned by different teams or vendors. Agents publish capabilities via Agent Cards and exchange work through stateful Tasks. Distinct from MCP, which governs agent-to-tool access, not agent-to-agent.
Agent Memory Decay
The gradual degradation of an agent's retrieved or stored context — stale reads, outdated embeddings, compounding errors carried forward from earlier steps — that surfaces as silent quality drift rather than a hard failure or a triggered alert.
Context Rot
The degradation of a model's ability to accurately recall and use information as the token count in its context window grows — even in models built for long context. Nothing errors; the model just gets quietly less accurate, which is what makes it easy to miss until output quality has already dropped.
State Handle Hijacking
An attack where a server-side state identifier (a handle) is obtained or guessed by an attacker and used to access another user's state on an MCP server. Introduced as a named attack surface by MCP's 2026-07-28 statelessness redesign, which removed session-based identity and made handle continuity the server's explicit responsibility instead of the protocol's.
Agent Harness
The software layer wrapped around a model that turns raw text generation into a working agent: the tool-execution loop, permission and approval gating, context and compaction management, and the guardrails that decide what the model's output is allowed to touch. The model generates text; the harness decides what that text can touch.
Need more than a definition?
If a term on this list describes a problem you have, talk to an engineer about it.