Glossary
Alphabetical definitions for the terms used across AGENT-33's documentation and source. Each entry includes a one-line definition and a pointer to where the concept is described in more depth.
A
Adapter. A class that translates between AGENT-33's internal
protocols and an external system. Examples: LLM provider adapters
(Ollama, OpenAI-compatible), messaging adapters (Telegram, Slack), code
execution adapters (CLI subprocess). See
docs/architecture/components.md.
Agent. A configured behavior — system prompt, allowed tools, model
assignment, iteration policy — defined as a JSON file under
engine/agent-definitions/. See docs/concepts.md.
Agent registry. The startup-time index of all loaded agent
definitions. Routes resolve agent names through the registry. See
docs/architecture/agents.md.
Agent runtime. The iterative or streaming loop that turns an agent
plus an input into an output. See docs/concepts.md.
Alembic. The database migration tool used for the PostgreSQL schema.
Migrations live in engine/alembic/versions/.
APScheduler. The cron-style scheduler that drives knowledge ingestion jobs.
Autonomy budget. A runtime envelope describing the file, command,
and network scope an agent is allowed to use, plus stop conditions. See
docs/concepts.md.
B
BM25. A lexical ranking function used as one half of the hybrid retrieval pipeline. The other half is vector similarity; they are combined through reciprocal rank fusion.
Browser agent. The reference agent that drives the headless browser tool. Useful for tasks that require navigating live web pages.
C
Candidate asset. An external resource (skill, agent, pack, knowledge
item) submitted to the platform but not yet published. Lives in a
governed lifecycle. See docs/concepts.md.
Checkpoint. A persisted record of a workflow's state after a step completes. Lets a crashed workflow resume from the last good point.
CLI. The agent33 command-line tool. Installable via
pip install -e ".[dev]". See docs/cli-reference.md.
Code execution. The subsystem (engine/src/agent33/execution/) that
runs code under a sandbox contract. Used by workflow steps that need to
execute generated code.
Conditional branch. A workflow step that selects one of several next
steps based on an expression. See
docs/architecture/workflows.md.
Confidence label. A tag on a candidate asset indicating how confident the submitter or the validator is in its quality. Values: low, medium, high.
D
DAG. Directed acyclic graph. The shape of every workflow.
DLQ. Dead-letter queue. Where messages go when a downstream handler
repeatedly fails to process them. See
engine/src/agent33/automation/.
E
Embedding. A vector representation of text used by the long-term memory and the hybrid retrieval pipeline.
Embedding provider. The adapter that produces embeddings. The active provider is chosen at startup by configuration.
Engine. The Python/FastAPI runtime. Lives under engine/.
Evaluation. The subsystem that runs golden tasks and golden cases, computes metrics, and detects regressions.
Expression evaluator. The small language used in workflow step
references — for example ${step.output.field}. See
docs/architecture/workflows.md.
F
Failure taxonomy. The 10-category classification applied to trace
failures so they can be grouped and reasoned about. See
docs/architecture/observability.md.
FastAPI. The Python web framework AGENT-33's engine is built on.
Frontend. The React/TypeScript operator console under frontend/.
G
Golden case. A specific test scenario used by the evaluation suite with known inputs and expected outputs.
Golden task. A high-level evaluation target made up of one or more golden cases.
Governance. The tool-policy layer that decides which tenants can call which tools under which conditions.
H
Headless browser. The browser automation tool that drives Chromium without a visible window.
Health check. A liveness probe on a subsystem. Each messaging adapter, each model provider, and each subsystem in the lifespan exposes a health check.
HPA. Horizontal Pod Autoscaler. The production Kubernetes overlay wires the engine deployment to one.
Hybrid search. Retrieval that combines BM25 (lexical) with vector similarity (semantic) via reciprocal rank fusion.
I
Ingestion. The candidate-asset lifecycle: submitted → triaged →
validating → published → revoked. See
engine/src/agent33/ingestion/.
Iterative invocation. Agent runtime mode where each step returns when the loop completes. The other mode is streaming.
J
JWT. JSON Web Token. One of the two supported authentication mechanisms. The other is API key.
K
Kernel container. A sandboxed Jupyter-style execution context.
Documented in
docs/runbooks/jupyter-kernel-containers.md.
Knowledge ingestion. The subsystem that pulls in external content from RSS, GitHub, web pages, and folders on a schedule.
L
L0 / L1 / L2. The three levels of progressive disclosure for skills: summary, outline, full body.
Lifespan. The FastAPI lifespan handler that initializes subsystems in order at startup and unwinds them at shutdown.
Lineage. Parent-child relationships between traces. Lets a single workflow's events be assembled into a tree.
LLM router. The layer that picks a provider + model for a given
agent call. See docs/architecture/components.md.
Long-term memory. The pgvector-backed store that holds embedded context across sessions, scoped to the tenant.
M
Manifest. A pack's declaration of its name, version, contents, and dependencies.
MCP. Model Context Protocol. AGENT-33 integrates as both a server
(exposing its surface) and a client (consuming external MCP servers).
See docs/architecture/mcp-integration.md.
Messaging adapter. A driver for a chat platform (Telegram, Discord, Slack, WhatsApp).
Metric. A measurable value emitted by the engine. Metric names
follow the agent33_* prefix.
Model router. Same as LLM router.
mypy. The static type checker. Runs in strict mode in CI.
N
NATS. The lightweight event bus used for asynchronous communication between subsystems.
O
Ollama. A local LLM runtime. The default provider in the Docker Compose stack.
Outcome. A recorded result of an agent or workflow run, used by the
impact dashboard and regression detection. See
docs/concepts.md.
Override. An explicit, audited deviation from a policy. Overrides leave a record in the audit trail.
P
Pack. A distributable bundle of skills, agents, tools, and policy
with a manifest and an integrity hash. See docs/concepts.md.
PackHub. The optional remote registry for packs.
Parallel group. A workflow step that runs multiple child steps in parallel and joins their results.
pgvector. The PostgreSQL extension that backs long-term memory.
Preflight check. An autonomy budget check applied before the agent runs, as opposed to the runtime enforcement that happens during execution.
Progressive disclosure. The L0/L1/L2 mechanism for showing the agent only as much skill content as it needs.
Provider catalog. The auto-registered list of available LLM providers, populated from environment variables at startup.
Q
QA agent. The reference agent that reviews output for correctness.
R
RAG. Retrieval-augmented generation. The pipeline that pulls relevant memory into the agent's prompt before the LLM call.
Redis. The in-memory store used for ephemeral state.
Registry. A startup-time index. Multiple registries exist: agent registry, skill registry, tool registry, pack registry, provider registry.
Release lifecycle. The state machine that governs release artifacts: planned → frozen → rc → validating → released → rolled_back.
Replay. The ability to re-execute a past run from its trace stream.
Reciprocal rank fusion (RRF). The algorithm that combines BM25 and vector similarity scores into a single ranking.
Retention policy. A rule that decides how long traces and outcomes are kept before being summarized and eventually purged.
Rollback. The release lifecycle transition that returns the system to a previous release artifact.
ruff. The Python linter and formatter. Replaces black, isort, and flake8 for this project.
S
Sandbox contract. The structured input to the code-execution subsystem describing what the executed code is allowed to do.
Session. A conversational unit of state — tenant, model, short-term memory, long-term memory scope, trace stream.
Session summarizer. The component that compresses short-term memory when it grows past a threshold.
Skill. A documented capability (Markdown or YAML) with frontmatter and a body. Participates in progressive disclosure.
Skill injector. The component that resolves which skills to include in a given agent invocation and at which level (L0/L1/L2).
SkillsBench. The third-party benchmark AGENT-33 runs to measure skill-driven capability. Smoke tier runs in CI; full tier runs weekly.
SSE. Server-sent events. The streaming transport for agent transcripts.
Stop condition. An autonomy-budget rule that ends an agent run when hit (max steps, max wall clock, max cost).
Streaming invocation. Agent runtime mode where events are yielded as they happen. Used by the operator console for live transcripts.
structlog. The Python structured logging library used throughout the engine.
T
Tenant. An isolated namespace. Every piece of state is scoped to a
tenant. See docs/concepts.md.
Tool. A function an agent can call. Validated by JSON Schema at
registration and at invocation. See docs/concepts.md.
Tool governance. The policy layer that decides which tenants can call which tools.
Topological sort. The algorithm used to order workflow steps for execution.
Trace. An audit record of an execution. See
docs/concepts.md.
Trust label. A tag on a candidate asset indicating its source: untrusted, community, maintainer, first-party.
U
Upsert. Insert-or-replace. The persistence pattern used by the ingestion subsystem and several others.
V
Vector store. The pgvector table that holds embedded long-term memory.
W
Workflow. A DAG of steps. See docs/concepts.md.
Workflow bridge. The component that lets a workflow step invoke an agent through the registry.
See also
docs/concepts.md— extended explanations of these termsARCHITECTURE.md— system overview