Configuration reference

Every AGENT-33 setting is exposed as an environment variable. They are loaded by agent33.config.Settings (pydantic-settings), so the variable name is the uppercased version of the field name in engine/src/agent33/config.py.

This page lists every setting, grouped by subsystem, with type, default, purpose, and an example. For how to organize configuration across environments, see setup-guide.md.

Loading priority

Values passed directly to Settings(...).
Environment variables.
.env then .env.local in the working directory.
Profile preset (AGENT33_PROFILE env var).
File secrets (mounted Kubernetes secrets).
Defaults baked into Settings.

Profiles

AGENT33_PROFILE selects a curated preset:

Profile	Intent
`minimal`	smoke tests, demos, CI
`developer`	local development
`production`	real deployments
`enterprise`	multi-tenant production with telemetry
`airgapped`	disconnected environments

Settings below show the default for agent33_mode=standard; a profile may change them.

API

Variable	Type	Default	Description
`API_PORT`	int	`8000`	TCP port the FastAPI server binds.
`API_LOG_LEVEL`	str	`info`	Uvicorn log level (`critical`, `error`, `warning`, `info`, `debug`, `trace`).
`API_SECRET_KEY`	secret	`change-me-in-production`	Internal HMAC / signing key; rotate in prod.
`CORS_ALLOWED_ORIGINS`	csv	`""` (deny all)	Comma-separated allowed origins. Empty means the browser cannot call the API.
`MAX_REQUEST_SIZE_BYTES`	int	`10485760`	Per-request body cap (10 MB).
`ENVIRONMENT`	str	`development`	One of `development`, `test`, `production`. Drives strict validators.

Example:

API_PORT=8000
API_LOG_LEVEL=info
API_SECRET_KEY=base64-random
CORS_ALLOWED_ORIGINS=https://control.example.com,https://agent33.example.com
ENVIRONMENT=production

Deployment mode

Variable	Type	Default	Description
`AGENT33_MODE`	enum	`standard`	One of `lite`, `standard`, `enterprise`.
`AGENT33_PROFILE`	str	(unset)	Loads a curated preset on top of env vars.
`OFFLINE_MODE`	bool	`false`	Disables outbound calls (pack registry, telemetry, remote LLMs).

Authentication and authorization

Variable	Type	Default	Description
`JWT_SECRET`	secret	`change-me-in-production`	HMAC secret for JWT signing. Required in production.
`JWT_ALGORITHM`	str	`HS256`	Token algorithm.
`JWT_EXPIRE_MINUTES`	int	`60`	Token lifetime.
`ENCRYPTION_KEY`	secret	`""`	Fernet key for encrypted persistence; required in `enterprise` mode.
`AUTH_BOOTSTRAP_ENABLED`	bool	`false`	Create a seed admin user on first start.
`AUTH_BOOTSTRAP_ADMIN_USERNAME`	str	`admin`	Bootstrap username.
`AUTH_BOOTSTRAP_ADMIN_PASSWORD`	secret	`""`	Bootstrap password (set to something strong).
`AUTH_BOOTSTRAP_ADMIN_SCOPES`	csv	(admin + base scopes)	Scopes granted to the bootstrap admin.

The startup validator refuses to boot in environment=production if JWT_SECRET is at the default. In development / lite mode, a random secret is generated and a warning logged.

Database (Postgres + pgvector)

Variable	Type	Default	Description
`DATABASE_URL`	str	`postgresql+asyncpg://agent33:agent33@postgres:5432/agent33`	Async SQLAlchemy URL.
`DB_POOL_SIZE`	int	`10`	Connection pool size per process.
`DB_MAX_OVERFLOW`	int	`20`	Overflow connections.
`DB_POOL_PRE_PING`	bool	`true`	Validate connections before use.
`DB_POOL_RECYCLE`	int	`1800`	Recycle connections after N seconds.

Postgres must be 14+ with the pgvector extension. The engine creates the extension on startup when it has rights.

Redis

Variable	Type	Default	Description
`REDIS_URL`	str	`redis://redis:6379/0`	Connection string.
`REDIS_MAX_CONNECTIONS`	int	`50`	Pool ceiling.

NATS

Variable	Type	Default	Description
`NATS_URL`	str	`nats://nats:4222`	Connection string.

JetStream must be enabled on the server.

LLM providers

Ollama (default local)

Variable	Type	Default	Description
`OLLAMA_BASE_URL`	str	`http://ollama:11434`	Base URL.
`OLLAMA_DEFAULT_MODEL`	str	`llama3.2:3b`	Model used when no model is requested.

LM Studio

Variable	Type	Default	Description
`LM_STUDIO_BASE_URL`	str	`http://localhost:1234/v1`	OpenAI-compatible LM Studio URL.
`LM_STUDIO_DEFAULT_MODEL`	str	`local-model`	Model identifier.

Local orchestration (llama.cpp tensor-offload)

Variable	Type	Default	Description
`LOCAL_ORCHESTRATION_MODEL`	str	`qwen3-coder-next`	Headline model.
`LOCAL_ORCHESTRATION_FORMAT`	str	`gguf_q4_k_m`	GGUF quantization.
`LOCAL_ORCHESTRATION_ENGINE`	str	`llama.cpp`	Engine name.
`LOCAL_ORCHESTRATION_BASE_URL`	str	`http://host.docker.internal:8033/v1`	OpenAI-compatible endpoint.

Cloud / OpenAI-compatible

Variable	Type	Default	Description
`DEFAULT_MODEL`	str	`""`	Override the default model. Use a provider prefix (e.g., `openai/gpt-4o-mini`, `openrouter/auto`).
`OPENAI_API_KEY`	secret	`""`	Enable OpenAI / compatible.
`OPENAI_BASE_URL`	str	`""` (defaults to OpenAI)	Custom base URL.
`OPENROUTER_API_KEY`	secret	`""`	OpenRouter key.
`OPENROUTER_BASE_URL`	str	`https://openrouter.ai/api/v1`
`OPENROUTER_SITE_URL`	str	`http://localhost`	Identifier for OpenRouter.
`OPENROUTER_APP_NAME`	str	`AGENT-33`
`OPENROUTER_APP_CATEGORY`	str	`cli-agent`
`OPENROUTER_DEFAULT_FALLBACK_MODELS`	csv	`""`	Comma-separated fallback chain.

AirLLM (GPU layer-sharded inference)

Variable	Type	Default	Description
`AIRLLM_ENABLED`	bool	`false`	Enable AirLLM provider.
`AIRLLM_MODEL_PATH`	str	`""`	Path or HF identifier.
`AIRLLM_DEVICE`	str	`cuda:0`	Device.
`AIRLLM_COMPRESSION`	str	`""`	`4bit` / `8bit` / empty.
`AIRLLM_MAX_SEQ_LEN`	int	`2048`
`AIRLLM_PREFETCH`	bool	`true`

Embeddings and retrieval

Variable	Type	Default	Description
`EMBEDDING_PROVIDER`	str	`ollama`	`ollama` or `jina`.
`EMBEDDING_DIM`	int	`768`	Must match the model output.
`EMBEDDING_BATCH_SIZE`	int	`100`
`EMBEDDING_CACHE_ENABLED`	bool	`true`
`EMBEDDING_CACHE_MAX_SIZE`	int	`1024`
`EMBEDDING_HOT_SWAP_ENABLED`	bool	`false`
`EMBEDDING_DEFAULT_MODEL`	str	`nomic-embed-text`
`EMBEDDING_DEFAULT_DIMENSIONS`	int	`768`
`EMBEDDING_QUANTIZATION_ENABLED`	bool	`false`	TurboQuant-style compression.
`EMBEDDING_QUANTIZATION_BITS`	int	`4`
`EMBEDDING_QUANTIZATION_SEED`	int	`42`
`RAG_TOP_K`	int	`5`
`RAG_SIMILARITY_THRESHOLD`	float	`0.3`
`RAG_HYBRID_ENABLED`	bool	`true`	BM25 + dense via RRF.
`RAG_VECTOR_WEIGHT`	float	`0.7`	BM25 weight = `1 - vector_weight`.
`RAG_RRF_K`	int	`60`	Reciprocal rank fusion constant.
`CHUNK_TOKENS`	int	`1200`
`CHUNK_OVERLAP_TOKENS`	int	`100`
`BM25_WARMUP_ENABLED`	bool	`true`
`BM25_WARMUP_MAX_RECORDS`	int	`10000`
`BM25_WARMUP_PAGE_SIZE`	int	`200`
`JINA_API_KEY`	secret	`""`
`JINA_READER_URL`	str	`https://r.jina.ai`

HTTP and runtime

Variable	Type	Default	Description
`HTTP_MAX_CONNECTIONS`	int	`20`	Per-client HTTP pool.
`HTTP_MAX_KEEPALIVE`	int	`10`

Policy / governance

Variable	Type	Default	Description
`TOOL_USE_MODE`	enum	`audit`	`audit`, `dry-run`, `approved`.
`EVIDENCE_REQUIRED`	bool	`true`	Reject runs without evidence capture.
`REVIEW_AUTHORITY`	enum	`user`	`user`, `automation`, `disabled`.
`REDACT_SECRETS_ENABLED`	bool	`true`	Redact API keys / tokens in logs and tool output.

Rate limiting

Variable	Type	Default
`RATE_LIMIT_ENABLED`	bool	`true`
`RATE_LIMIT_DEFAULT_TIER`	str	`standard`
`RATE_LIMIT_PER_MINUTE`	int	`60`
`RATE_LIMIT_BURST`	int	`10`

Web search

Variable	Type	Default	Description
`SEARXNG_URL`	str	`http://searxng:8080`
`TAVILY_API_KEY`	secret	`""`
`BRAVE_API_KEY`	secret	`""`
`WEB_SEARCH_MAX_RESULTS`	int	`10`
`WEB_SEARCH_DEFAULT_PROVIDER`	str	(unset)	Force a provider; empty auto-selects.

Voice

Variable	Type	Default	Description
`VOICE_DAEMON_ENABLED`	bool	`true`
`VOICE_DAEMON_TRANSPORT`	enum	`stub`	`stub`, `sidecar`, `livekit`.
`VOICE_DAEMON_URL`	str	`""`	Sidecar URL.
`VOICE_DAEMON_API_KEY`	secret	`""`
`VOICE_DAEMON_API_SECRET`	secret	`""`
`VOICE_DAEMON_ROOM_PREFIX`	str	`agent33-voice`
`VOICE_DAEMON_MAX_SESSIONS`	int	`25`
`VOICE_TTS_PROVIDER`	enum	`stub`	`stub`, `elevenlabs`, `piper`.
`VOICE_STT_PROVIDER`	enum	`stub`	`stub`, `whisper`, `openai_whisper`.
`VOICE_STT_WHISPER_MODEL_SIZE`	str	`base`	tiny/base/small/medium/large.
`VOICE_STT_WHISPER_DEVICE`	str	`cpu`	`cpu` / `cuda`.
`VOICE_TTS_PIPER_MODEL_PATH`	str	`""`
`VOICE_TTS_PIPER_VOICE_ID`	str	`en_US-lessac-medium`
`ELEVENLABS_API_KEY`	secret	`""`
`ELEVENLABS_VOICE_ID`	str	`21m00Tcm4TlvDq8ikWAM`
`VOICE_ELEVENLABS_ENABLED`	bool	`false`
`VOICE_ELEVENLABS_API_KEY`	secret	`""`
`VOICE_ELEVENLABS_DEFAULT_VOICE_ID`	str	`""`
`VOICE_ELEVENLABS_MODEL_ID`	str	`eleven_multilingual_v2`
`VOICE_LIVEKIT_ENABLED`	bool	`false`
`VOICE_LIVEKIT_API_KEY`	secret	`""`
`VOICE_LIVEKIT_API_SECRET`	secret	`""`
`VOICE_LIVEKIT_WS_URL`	str	`""`

Knowledge ingestion

Variable	Type	Default	Description
`KNOWLEDGE_DEFAULT_TENANT_ID`	str	`system`
`INGESTION_DB_PATH`	str	`var/ingestion.db`
`INGESTION_MAILBOX_DB_PATH`	str	`var/ingestion_mailbox.db`
`INGESTION_JOURNAL_DB_PATH`	str	`var/ingestion_journal.db`
`INGESTION_JOURNAL_RETENTION_DAYS`	int	`90`	0 disables expiry.
`INGESTION_TASK_METRICS_DB_PATH`	str	`var/ingestion_task_metrics.db`
`INGESTION_TASK_METRICS_RETENTION_DAYS`	int	`30`
`INGESTION_NOTIFICATION_HOOKS_DB_PATH`	str	`var/ingestion_notification_hooks.db`
`INGESTION_NOTIFICATION_TIMEOUT_SECONDS`	float	`5.0`

Packs

Variable	Type	Default	Description
`PACK_DEFINITIONS_DIR`	str	`packs`
`PACK_MARKETPLACE_DIR`	str	`pack-marketplace`
`PACK_MARKETPLACE_REMOTE_SOURCES`	json	`""`	JSON array of remote source configs.
`PACK_MARKETPLACE_CACHE_DIR`	str	`var/pack-marketplace-cache`
`PACK_AUTO_ENABLE`	bool	`false`
`PACK_MAX_SIZE_MB`	int	`50`
`PACK_CHECKSUMS_REQUIRED`	bool	`false`
`PACK_ROLLBACK_ARCHIVE_DIR`	str	`var/pack-rollback-archive`
`PACK_CURATION_ENABLED`	bool	`false`
`PACK_MIN_QUALITY_SCORE`	float	`0.5`
`PACK_REQUIRE_REVIEW_FOR_LISTING`	bool	`true`
`PACK_DEFAULT_CATEGORIES`	csv	(automation, data-analysis, ...)
`PACK_SIGNING_KEY`	str	`""`

Skills

Variable	Type	Default
`SKILL_DEFINITIONS_DIR`	str	`skills`
`SKILL_MAX_INSTRUCTIONS_CHARS`	int	`16000`
`SKILL_LINEAGE_STORE_PATH`	str	`var/skill_lineage_events.json`
`SKILLSBENCH_SKILL_MATCHER_ENABLED`	bool	`false`
`SKILLSBENCH_SKILL_MATCHER_MODEL`	str	`llama3.2`
`SKILLSBENCH_SKILL_MATCHER_TOP_K`	int	`20`
`SKILLSBENCH_SKILL_MATCHER_SKIP_LLM_BELOW`	int	`3`
`SKILLSBENCH_CONTEXT_MANAGER_ENABLED`	bool	`true`
`SKILLSBENCH_STORAGE_PATH`	str	`var/skillsbench_runs`
`SKILL_MATCH_FUZZY_THRESHOLD`	float	`0.7`
`SKILL_MATCH_SEMANTIC_THRESHOLD`	float	`0.5`
`SKILL_MATCH_CONTEXTUAL_THRESHOLD`	float	`0.4`
`SKILL_MATCH_MAX_CANDIDATES`	int	`10`

Plugins

Variable	Type	Default
`PLUGIN_DEFINITIONS_DIR`	str	`plugins`
`PLUGIN_AUTO_ENABLE`	bool	`true`
`PLUGIN_STATE_STORE_PATH`	str	`var/plugin_lifecycle_state.json`
`PLUGIN_ALLOWLIST`	csv	`""` (allow all)
`PLUGIN_DISCOVERY_PATHS`	csv	`""`

Agents

Variable	Type	Default	Description
`AGENT_DEFINITIONS_DIR`	str	`agent-definitions`
`AGENT_EFFORT_ROUTING_ENABLED`	bool	`false`	Enable effort-based model routing.
`AGENT_EFFORT_DEFAULT`	str	`medium`
`AGENT_EFFORT_LOW_MODEL`	str	`""`	Model for low effort.
`AGENT_EFFORT_MEDIUM_MODEL`	str	`""`
`AGENT_EFFORT_HIGH_MODEL`	str	`""`
`AGENT_EFFORT_LOW_TOKEN_MULTIPLIER`	float	`1.0`
`AGENT_EFFORT_MEDIUM_TOKEN_MULTIPLIER`	float	`1.0`
`AGENT_EFFORT_HIGH_TOKEN_MULTIPLIER`	float	`1.0`
`AGENT_EFFORT_HEURISTIC_ENABLED`	bool	`true`
`AGENT_EFFORT_POLICY_TENANT`	json	`""`	`{"tenant-id": "low
`AGENT_EFFORT_POLICY_DOMAIN`	json	`""`
`AGENT_EFFORT_POLICY_TENANT_DOMAIN`	json	`""`
`AGENT_EFFORT_COST_PER_1K_TOKENS`	float	`0.0`	Cost basis for estimates.
`AGENT_DEFAULT_CONTEXT_WINDOW`	int	`128000`
`AGENT_CONTEXT_WARN_THRESHOLD`	float	`0.8`
`AGENT_TOOL_LOOP_MAX_RETRIES`	int	`3`
`AGENT_TOOL_LOOP_BACKOFF_BASE_MS`	float	`100`
`AGENT_PROFILER_MAX_PROFILES`	int	`1000`

Autonomy

Variable	Type	Default	Description
`AUTONOMY_DEFAULT_LEVEL`	int	`1`	`0=supervised`, `1=read-auto`, `2=auto-no-destructive`, `3=full`.
`AUTONOMY_MAX_STRETCH_HOURS`	int	`24`
`AUTONOMY_ALLOW_SECURITY_RECON`	bool	`true`

Code execution

Variable	Type	Default	Description
`EXECUTION_GPU_ENABLED`	bool	`false`
`EXECUTION_DEFAULT_DOCKER_IMAGE`	str	`python:3.11-slim`
`EXECUTION_GPU_RUNTIME`	enum	`nvidia`	`nvidia` or `amd`.

Jupyter kernel

Variable	Type	Default	Description
`JUPYTER_KERNEL_ENABLED`	bool	`false`
`JUPYTER_KERNEL_ADAPTER_ID`	str	`jupyter-kernel`
`JUPYTER_KERNEL_TOOL_ID`	str	`code-interpreter`
`JUPYTER_KERNEL_MODE`	enum	`local`	`local` or `docker`.
`JUPYTER_KERNEL_NAME`	str	`python3`
`JUPYTER_KERNEL_MAX_SESSIONS`	int	`10`
`JUPYTER_KERNEL_IDLE_TIMEOUT_SECONDS`	float	`300.0`
`JUPYTER_KERNEL_STARTUP_TIMEOUT_SECONDS`	float	`30.0`
`JUPYTER_KERNEL_EXECUTION_TIMEOUT_SECONDS`	float	`60.0`
`JUPYTER_KERNEL_DOCKER_IMAGE`	str	`quay.io/jupyter/minimal-notebook:python-3.11`
`JUPYTER_KERNEL_ALLOWED_IMAGES`	csv	`""`
`JUPYTER_KERNEL_NETWORK_ENABLED`	bool	`false`
`JUPYTER_KERNEL_MOUNT_WORKDIR`	bool	`true`
`JUPYTER_KERNEL_CONTAINER_WORKDIR`	str	`/workspace`

Hooks

Variable	Type	Default
`HOOKS_ENABLED`	bool	`true`
`HOOKS_DEFINITIONS_DIR`	str	`hook-definitions`
`HOOKS_DEFAULT_TIMEOUT_MS`	float	`200.0`
`HOOKS_CHAIN_TIMEOUT_MS`	float	`500.0`
`HOOKS_FAIL_OPEN_DEFAULT`	bool	`true`
`HOOKS_MAX_PER_EVENT`	int	`20`
`HOOKS_EXECUTION_LOG_ENABLED`	bool	`true`
`HOOKS_EXECUTION_LOG_RETENTION_HOURS`	int	`24`
`SCRIPT_HOOKS_ENABLED`	bool	`true`
`SCRIPT_HOOKS_PROJECT_DIR`	str	(defaults to `<cwd>/.claude/hooks/`)
`SCRIPT_HOOKS_USER_DIR`	str	(defaults to `~/.agent33/hooks/`)
`SCRIPT_HOOKS_DEFAULT_TIMEOUT_MS`	float	`5000.0`
`SCRIPT_HOOKS_MAX_TIMEOUT_MS`	float	`30000.0`

MCP (Model Context Protocol)

Variable	Type	Default	Description
`MCP_SERVERS`	csv	`""`	Comma-separated MCP server URLs.
`MCP_TIMEOUT_SECONDS`	float	`30.0`
`MCP_AUTO_DISCOVER`	bool	`true`
`MCP_PROXY_CONFIG_PATH`	str	`""`
`MCP_PROXY_ENABLED`	bool	`false`
`MCP_PROXY_TOOL_SEPARATOR`	str	`__`
`MCP_PROXY_HEALTH_CHECK_ENABLED`	bool	`true`
`MCP_SYNC_BACKUP_ENABLED`	bool	`true`

Tool approvals

Variable	Type	Default	Description
`APPROVAL_TOKEN_TTL_SECONDS`	int	`300`
`APPROVAL_TOKEN_ENABLED`	bool	`true`
`APPROVAL_TOKEN_ONE_TIME_DEFAULT`	bool	`true`
`P69B_DB_PATH`	str	`var/p69b.db`
`TOOL_DISCOVERY_MODE`	enum	`legacy`	`legacy` or `dynamic`.

Connector boundary

Variable	Type	Default
`CONNECTOR_BOUNDARY_ENABLED`	bool	`false`
`CONNECTOR_POLICY_PACK`	str	`default`
`CONNECTOR_GOVERNANCE_BLOCKED_CONNECTORS`	csv	`""`
`CONNECTOR_GOVERNANCE_BLOCKED_OPERATIONS`	csv	`""`
`CONNECTOR_CIRCUIT_BREAKER_ENABLED`	bool	`false`
`CONNECTOR_CIRCUIT_FAILURE_THRESHOLD`	int	`3`
`CONNECTOR_CIRCUIT_RECOVERY_SECONDS`	float	`30.0`
`CONNECTOR_CIRCUIT_HALF_OPEN_SUCCESSES`	int	`2`
`CONNECTOR_CIRCUIT_MAX_RECOVERY_SECONDS`	float	`300.0`

Observability

Variable	Type	Default
`METRICS_ROLLING_WINDOW_SECONDS`	int	`300`
`OBSERVABILITY_EFFORT_ALERTS_ENABLED`	bool	`true`
`OBSERVABILITY_EFFORT_ALERT_HIGH_EFFORT_COUNT_THRESHOLD`	int	`25`
`OBSERVABILITY_EFFORT_ALERT_HIGH_COST_USD_THRESHOLD`	float	`5.0`
`OBSERVABILITY_EFFORT_ALERT_HIGH_TOKEN_BUDGET_THRESHOLD`	int	`8000`
`OBSERVABILITY_EFFORT_EXPORT_ENABLED`	bool	`false`
`OBSERVABILITY_EFFORT_EXPORT_PATH`	str	`var/effort_routing_events.jsonl`
`OBSERVABILITY_EFFORT_EXPORT_FAIL_CLOSED`	bool	`false`
`PROVENANCE_ENABLED`	bool	`true`
`PROVENANCE_MAX_RECEIPTS`	int	`10000`
`SLO_AVAILABILITY_TARGET`	float	`0.999`
`SLO_LATENCY_P99_MS`	int	`500`
`SLO_LATENCY_AGENT_P99_MS`	int	`10000`
`SLOW_QUERY_THRESHOLD_MS`	int	`100`

Streaming

Variable	Type	Default	Description
`STREAMING_MAX_CONNECTIONS`	int	`100`
`STREAMING_PING_INTERVAL_SECONDS`	int	`30`
`WORKFLOW_TRANSPORT_PREFERRED`	enum	`auto`	`auto`, `websocket`, or `sse`.
`WORKFLOW_WS_PING_INTERVAL`	float	`30.0`
`WORKFLOW_WS_PING_TIMEOUT`	float	`10.0`
`SSE_SCHEMA_V2_ENABLED`	bool	`false`	Opt-in to SSE v2 schema.

Browser automation

Variable	Type	Default	Description
`BROWSER_CLOUD_API_KEY`	secret	`""`	BrowserBase key; empty = local only.
`BROWSER_SESSION_TTL_SECONDS`	int	`300`
`BROWSER_VISION_MODEL`	str	`""`
`BROWSER_CLOUD_API_URL`	str	`https://www.browserbase.com/v1`
`BROWSER_COMPUTER_USE_ENABLED`	bool	`false`
`MAX_BROWSER_SESSIONS_PER_TENANT`	int	`3`

Messaging integrations (Matrix)

Variable	Type	Default
`MATRIX_HOMESERVER_URL`	str	`""`
`MATRIX_ACCESS_TOKEN`	secret	`""`
`MATRIX_USER_ID`	str	`""`
`MATRIX_ROOM_IDS`	csv	`""` (empty = all joined)
`MATRIX_SYNC_TIMEOUT_MS`	int	`30000`

Self-improvement

Variable	Type	Default	Description
`SELF_IMPROVE_ENABLED`	bool	`true`
`SELF_IMPROVE_SCOPE`	csv	`prompts,workflows,templates`
`SELF_IMPROVE_REQUIRE_APPROVAL`	bool	`true`
`SELF_IMPROVE_PROPOSAL_SANDBOX_ENABLED`	bool	`true`
`IMPROVEMENT_LEARNING_ENABLED`	bool	`false`
`IMPROVEMENT_TUNING_LOOP_ENABLED`	bool	`false`
`IMPROVEMENT_TUNING_LOOP_INTERVAL_HOURS`	float	`24.0`
`IMPROVEMENT_TUNING_LOOP_REQUIRE_APPROVAL`	bool	`true`
`IMPROVEMENT_TUNING_LOOP_DRY_RUN`	bool	`false`
`IMPROVEMENT_LEARNING_PERSISTENCE_BACKEND`	enum	`db`	`memory`, `file`, `db`.

The improvement subsystem has more fine-grained knobs (quality weights, thresholds, retention); see engine/src/agent33/config.py for the full list.

Outcomes and impact

Variable	Type	Default
`OUTCOMES_DB_PATH`	str	`var/outcomes.db`
`PPACK_V3_ENABLED`	bool	`false`
`PPACK_V3_AB_ENABLED`	bool	`true`
`PPACK_V3_AB_EXPERIMENT_KEY`	str	`ppack_v3`
`PPACK_V3_AB_MIN_SAMPLES_PER_VARIANT`	int	`30`
`PPACK_V3_AB_REGRESSION_THRESHOLD`	float	`-0.05`
`PPACK_V3_AB_WEEKLY_WINDOW_DAYS`	int	`7`

Evaluation

Variable	Type	Default	Description
`EVALUATION_JUDGE_MODEL`	str	`""`	Empty = rule-based evaluator.
`EVALUATION_CTRF_OUTPUT_DIR`	str	`var/ctrf-reports`
`EVALUATION_BENCHMARK_CATALOG_PATH`	str	`""`
`EVALUATION_BENCHMARK_DEFAULT_TRIALS`	int	`5`
`SCHEDULED_GATES_ENABLED`	bool	`false`
`SCHEDULED_GATES_MAX_SCHEDULES`	int	`50`
`SCHEDULED_GATES_HISTORY_RETENTION`	int	`100`

Comparative scoring

Variable	Type	Default
`COMPARATIVE_ELO_K_FACTOR`	float	`32.0`
`COMPARATIVE_MIN_POPULATION_SIZE`	int	`2`
`COMPARATIVE_CONFIDENCE_LEVEL`	float	`0.95`

Operator sessions

Variable	Type	Default
`OPERATOR_SESSION_ENABLED`	bool	`true`
`OPERATOR_SESSION_BASE_DIR`	str	(`~/.agent33/sessions/`)
`OPERATOR_SESSION_CHECKPOINT_INTERVAL_SECONDS`	float	`60.0`
`OPERATOR_SESSION_MAX_REPLAY_FILE_MB`	int	`50`
`OPERATOR_SESSION_MAX_RETAINED`	int	`100`
`OPERATOR_SESSION_CRASH_RECOVERY_ENABLED`	bool	`true`

Context engine

Variable	Type	Default
`PROMPT_CACHE_ENABLED`	bool	`true`
`CONTEXT_COMPRESSION_ENABLED`	bool	`false`
`CONTEXT_COMPRESSION_THRESHOLD_PERCENT`	float	`0.50`
`CONTEXT_COMPRESSION_PROTECT_FIRST_N`	int	`3`
`CONTEXT_COMPRESSION_TAIL_TOKEN_BUDGET`	int	`20000`
`CONTEXT_COMPRESSION_SUMMARY_TARGET_RATIO`	float	`0.20`
`CONTEXT_COMPRESSION_SUMMARY_TOKENS_CEILING`	int	`12000`
`CONTEXT_COMPRESSION_SUMMARIZE_MODEL`	str	`llama3.2`
`CONTEXT_ENGINE_DEFAULT`	str	`builtin`
`CONTEXT_COMPACTION_ENABLED`	bool	`true`
`SESSION_SPAWN_TEMPLATES_DIR`	str	`""`
`SESSION_ARCHIVE_RETENTION_DAYS`	int	`90`

Webhook delivery

Variable	Type	Default
`WEBHOOK_DELIVERY_MAX_RETRIES`	int	`5`
`WEBHOOK_DELIVERY_BASE_DELAY`	float	`1.0`
`WEBHOOK_DELIVERY_MAX_DELAY`	float	`300.0`
`WEBHOOK_DELIVERY_MAX_RECORDS`	int	`10000`

Workflows and runs

Variable	Type	Default
`WORKFLOW_MARKETPLACE_ENABLED`	bool	`true`
`WORKFLOW_TEMPLATES_DIR`	str	`workflow-templates`
`SYNTHETIC_ENV_WORKFLOW_DIR`	str	`workflow-definitions`
`SYNTHETIC_ENV_TOOL_DIR`	str	`tool-definitions`
`SYNTHETIC_ENV_BUNDLE_RETENTION`	int	`100`
`SYNTHETIC_ENV_BUNDLE_PERSISTENCE_PATH`	str	`var/synthetic_environment_bundles.json`
`WORKFLOW_RUN_ARCHIVE_DIR`	str	`var/workflow-runs`
`ORCHESTRATION_STATE_STORE_PATH`	str	`var/orchestration_state.json`
`CHECKPOINT_PERSISTENCE_ENABLED`	bool	`false`

Process registry

Variable	Type	Default	Description
`PROCESS_MANAGER_LOG_DIR`	str	`var/process-manager`
`PROCESS_MANAGER_MAX_PROCESSES`	int	`10`

Backups

Variable	Type	Default	Description
`BACKUP_DIR`	str	`var/backups`

Trajectory capture and MoA

Variable	Type	Default
`TRAJECTORY_CAPTURE_ENABLED`	bool	`false`
`TRAJECTORY_OUTPUT_DIR`	str	`trajectories`
`MOA_REFERENCE_MODELS`	csv	`""`
`MOA_AGGREGATOR_MODEL`	str	`""`
`MOA_REFERENCE_TEMPERATURE`	float	`0.6`
`MOA_AGGREGATOR_TEMPERATURE`	float	`0.4`

Programmatic tool calling (PTC)

Variable	Type	Default
`PTC_ENABLED`	bool	`true`
`PTC_TIMEOUT_S`	int	`300`
`PTC_MAX_CALLS`	int	`50`
`PTC_MAX_STDOUT_BYTES`	int	`51200`
`PTC_ALLOWED_TOOLS`	csv	`""`

Control plane backend

Variable	Type	Default	Description
`CONTROL_PLANE_BACKEND`	enum	`sqlite`	`memory` or `sqlite`.
`CONTROL_PLANE_DB_PATH`	str	`agent33_control_plane.db`

Training

Variable	Type	Default
`TRAINING_ENABLED`	bool	`true`
`TRAINING_OPTIMIZE_INTERVAL`	int	`100`
`TRAINING_IDLE_OPTIMIZE_SECONDS`	int	`300`
`TRAINING_MIN_ROLLOUTS`	int	`10`
`TRAINING_EVAL_MODEL`	str	`""`

Alembic

Variable	Type	Default	Description
`ALEMBIC_CONFIG_PATH`	str	`alembic.ini`
`ALEMBIC_AUTO_CHECK_ON_STARTUP`	bool	`false`

Component security scans

Variable	Type	Default
`COMPONENT_SECURITY_SCAN_STORE_ENABLED`	bool	`false`
`COMPONENT_SECURITY_SCAN_STORE_DB_PATH`	str	`var/component_security_scans.sqlite3`
`COMPONENT_SECURITY_SCAN_STORE_RETENTION_DAYS`	int	`90`

Lite mode

Variable	Type	Default	Description
`SQLITE_MEMORY_DB_PATH`	str	`var/agent33_memory.db`	Lite-mode long-term memory path. Use `:memory:` for ephemeral.

configuration