Configuration reference

Every tunable in one place, with defaults. Names link to the guide that explains them.

Engine and runs

Setting	Where	Default
store	Engine(...)	InMemoryEventStore()	Persistence
bus	Engine(...)	InProcessBus()	DistributedBus, ClientBus for clusters
schemas	Engine(...)	empty SchemaRegistry	versioned payload validation
default_budget	Engine(...)	unlimited	applies when a run has no budget=
blobs, claim_threshold	Engine(...)	none, 16384	Large payloads
sweep_interval_s	Engine(...)	none	periodic sweep()
store_retry	Engine(...)	RetryPolicy() (30 s, 0.1 s to 2 s backoff, jitter)	how long to ride out a store outage before the bus redelivers instead
is_transient	Engine(...)	aaa.core.is_transient	which exceptions count as an outage (vs a bug)
max_steps, max_tokens, max_cost, deadline_s	Budget(...)	none	per-run limits
run_id	engine.start/run(...)	generated ULID	makes starting idempotent
parent, tag	engine.start(...)	none	child-run linkage

LLM

Setting	Where	Default
provider, model, base_url, api_key, headers, chat_path, embeddings_path, timeout_s, params, price_in_per_mtok, price_out_per_mtok	LLMConfig	see LLM gateways
concurrency, emit_deltas	LLMHandler(...)	8, True
max_cached_conversations	LLMHandler attribute	1024
max_attempts, base_delay, max_delay	RetryingClient	3, 0.5, 20.0
max_attempts, max_elapsed_s, base_delay_s, max_delay_s, multiplier, jitter	RetryPolicy	None, 30.0, 0.1, 2.0, 2.0, True
failure_threshold, cooldown_s	CircuitBreakerClient	5, 30.0

Tools

Setting	Where	Default
requires_approval, timeout_s, pool, name, description	@tool(...)	False, 60, "default"
delivery, idempotency_key, guard	@tool(...)	handler default
pools	ToolHandler(...)	{"default": 16, "delegate": 8}
result_retry	ToolHandler(...)	RetryPolicy.forever() (never lose a result of a tool that already ran)
guard	ToolHandler(...)	GuardConfig()
enabled, delivery, max_attempts, lease_ttl_s, heartbeat_s, wait_poll_s, on_lost_lease, fence_results	GuardConfig	True, at_least_once, 3, 15.0, ttl/3, 0.5, cancel, True

Algorithms

Algorithm	Settings (defaults)
ToolLoop / ReAct	system="", model=None, max_turns=10, params={}, tool_names=None, name, stream=False
PlanAndExecute	model=None, max_parallel=3, max_replans=1, synthesize=True, planner_system
Router	classifier=None, default=None, model=None
Supervisor	workers, extra_tools=None, plus ToolLoop settings
GenerateTestRepair	validator="default", system="", model=None, max_iterations=3, params={}, extract_code=True
GeneratorCritic	critic_system, critic_model=None, plus the above
SelfConsistency	n=5, temperature=0.8, model=None, system, extract
RAGAgent	index, k=5, plus ToolLoop settings
Projection	name="", version=1, state_type=None, every=64

Stores and bus

Setting	Where	Default
path, synchronous, poll_s	SQLiteEventStore	required, "NORMAL", 0.2
dsn, partitions, outbox, max_pool	PostgresEventStore.create	required, 16, True, 10
url, partitions, prefix, outbox	RedisEventStore.create	required, 16, "aaa", True
member, batch, poll_s, lease_ttl_s, tick_s, handover_timeout_s, redeliver_base_s, redeliver_max_s	DistributedBus	random, 64, 0.5, 6.0, 1.0, 30.0, 0.2, 5.0
queue_size, redeliver_base_s, redeliver_max_s	InProcessBus	1024, 0.2, 5.0
keep_last	store.trim_outbox	100_000

Environment variables

AAA reads no environment variables itself, except where you reference them as ${VAR} in LLMConfig (api_key, headers). The gateway presets reference these names, which you set:

Preset	Variables
litellm_proxy	LITELLM_API_KEY
portkey	PORTKEY_API_KEY, PROVIDER_API_KEY
cloudflare	CLOUDFLARE_API_TOKEN, PROVIDER_API_KEY
vercel	AI_GATEWAY_API_KEY

For the test suite and live smoke test: AAA_TEST_POSTGRES_URL, AAA_TEST_REDIS_URL, AAA_TEST_PG_BIN, and AAA_LIVE_BASE_URL, AAA_LIVE_MODEL, AAA_LIVE_API_KEY, AAA_LIVE_HEADERS, AAA_LIVE_CHAT_PATH, AAA_LIVE_EMBED_MODEL, AAA_LIVE_TIMEOUT.