Benchmarks

Median time from guest message to agent-generated reply across all active venue agents. Includes LLM inference, tool execution, and message delivery.

Hard timeout for every tool invocation (venue search, booking, calorie scan, etc.). If a tool exceeds 5 seconds, it fails gracefully and the agent responds without it.

Maximum wait time for a ClawPulse A2A message round-trip. Includes envelope signing, network transit, venue agent processing, and response delivery.

Maximum allowed time for a single Claude API call. In practice, Haiku completes in 1-3s and Sonnet in 3-8s. The 45s cap handles edge cases with complex multi-tool conversations.

A2A gateway circuit breaker opens after 5 failures within a 10-minute window. Applied both globally and per-venue. Half-open probe allows one request to test recovery after 5-minute cooldown.

Failed message sends retry with exponential backoff: 1 minute, 5 minutes, 25 minutes. Permanent failures (invalid recipient, 404, 401) go directly to DLQ. Transient failures (timeout, 503) get all 3 attempts.

Per-platform circuit breaker for WhatsApp/Telegram/Instagram message delivery. Opens after 10 consecutive failures, resets after 2 minutes. Prevents cascading failure when a platform is down.

Permanently failed messages are stored in a dead letter queue capped at 5,000 entries with 7-day retention. Alerts fire when the queue exceeds 4,000 entries.

Claude Haiku 4.5 handles free and starter tier conversations. Output cost: $4.00/1M tokens. Max output: 1,024 tokens per response. Used for simple queries, FAQ responses, and venue search.

Claude Sonnet 4.6 handles pro tier conversations. Output cost: $15.00/1M tokens. Max output: 2,048 tokens per response. Used for complex multi-tool queries, booking coordination, and personalized recommendations.

Metered billing for external agents using the AGNT Open Network. Venue searches are free. Each booking.confirm intent costs $0.25. Batched nightly to Stripe usage records.

Hard daily cap on total LLM inference spending across all users. Prevents runaway costs. Tracked in Redis with 24-hour expiry. Fails open on Redis outage to avoid blocking users.

Each non-trivial user message is embedded and searched against stored memory via pgvector cosine distance. Top 10 relevant facts are injected into the system prompt alongside structural keys.

Up to 20 messages stored per conversation, 12 sent to the LLM per turn. Conversation TTL is 24 hours. History is encrypted at rest with Fernet and synced to PostgreSQL as backup.

Constructed soul prompts (structural memory + context) are cached in Redis for 1 hour. Cache is invalidated immediately on memory writes so the agent always has fresh context.

Maximum 20 concurrent embedding API calls with 30-second timeout per call. Prevents overwhelming the embedding service during high-traffic memory write bursts.

Global semaphore limiting concurrent Claude API calls to 30. Additional requests queue with a 15-second acquisition timeout. Prevents API rate limit exhaustion under load.

Daily token limit for free-tier users. Tracked in Redis with 24-hour expiry. At approximately 500 tokens per message, this supports about 100 conversations per day.

Daily token limit for pro-tier users and venue pro subscriptions. 20x the free tier. Supports sustained high-volume usage including multi-tool conversations and complex queries.

Default rate limit for general API endpoints. LLM-specific endpoints: 10/min. Booking endpoints: 5/min. Redis-backed with fail-open on Redis outage.