| Ecosystem | A multi-agent workflow declared in YAML. Compiled into one Step Functions state machine. |
| Agent | A single role in an ecosystem — a prompt, tools, an output_kind, and a place in the DAG. |
| Swarm | A named group of agents that can run in parallel. |
| DAG | The directed acyclic graph formed by agents' depends_on edges. |
| output_kind | The contract the runtime applies to an agent (analyst, synthesizer, challenger, reporter, drafter, creative, chat, operator). |
| output_format | An optional typed artifact an agent produces (markdown/json/yaml/html). |
| Memory namespace | A shared store agents read/write to pass context within a run. |
| Report agent | Auto-injected agent (when report: true) that assembles the final report. |
| Challenger agent | Auto-injected agent (when quality.challenger: true) that stress-tests the output. |
| Grounding | Checking claims against sources — internal (Bedrock Guardrail) and external (web search). |
| Domain | A knowledge scope defining what to ingest and how. |
| Catalog | The grounded, versioned knowledge store agents cite. |
| Ingest | The build pipeline that populates the catalog (make ingest). |
| Synthesis | A catalog stage where agents derive higher-order entities from ingested facts. |
| Quorum | A deliberation model where a panel of scrutineer personas reach a verdict. |
| Scrutineer | A persona (with a lens) used in quorum deliberation. |
| Mission | The operator's request for a run, supplied via the ecosystem's inputs. |
| Report | The final structured deliverable of a run. |
| AgentCore Runtime | The single Bedrock AgentCore container that runs every agent. |
| AgentCore Gateway | Exposes MCP tools to agents. |
| AgentCore Memory | The shared, namespace-isolated memory store. |
| Seed | make seed — publish content + compile ecosystems (fast). |
| Evaluation harness | Offline capability eval + regression deploy-gate + online monitoring for an ecosystem. |
| Golden set | Versioned, labelled test cases (cases.jsonl) with known-correct answers. |
| Scorecard | The aggregated result of an eval run (per-metric median ± band, gate decision, miss-rates). |
| Noise band | The variance envelope across replicates; a change is a regression only if it clears the band. |
| Deploy gate | gate: block | warn | off — what a failing eval threshold does at deploy. |
| Online eval | Async, zero-latency evaluation of live production runs for drift detection. |
| Eval State Machine | The per-ecosystem Step Functions machine compiled when an ecosystem declares evaluation:. |
| Bedrock Evaluations | LLM-as-judge scoring service used by the bedrock_judge scorer. |
| Step Functions | AWS orchestration; one state machine per ecosystem (plus catalog and eval machines). |
| Verified Permissions | AWS authorization service; per-ecosystem access via Cedar policies. |