Provon Capabilities

Provon is organized around the production loop for AI agents: route model calls, capture evidence, replay failures, evaluate behavior, control cost, and enforce guardrails close to the Gateway path.

Capability Map

Gateway

Use one OpenAI-compatible surface for model traffic.

Gateway capabilities include:

• Provider keys, credentials, model capabilities, and per-project routing.
• Fallback, retry, circuit breaker, load balancing, and health-aware routing.
• Cost-aware and budget-aware routing strategies.
• Request capture and OpenTelemetry trace generation.
• Project pause controls and API key capability checks.

Built-in provider families include OpenAI, Anthropic, Azure OpenAI, OpenRouter, Gemini, Workers AI, Cohere, Groq, Mistral, DeepSeek, Together AI, Fireworks AI, Cerebras, SambaNova, Kimi, Doubao, MiniMax, Qwen, Z.ai, xAI, Perplexity, StepFun, custom OpenAI-compatible endpoints, and local/private provider targets.

Use Gateway when you need:

• One API contract across multiple model backends.
• A safer place to enforce model routing and fallback decisions.
• Traceable model attempts with request, response, token, cost, and error evidence.

Observability

Provon stores agent evidence as queryable telemetry.

Observability capabilities include:

• OTLP HTTP ingest for traces, logs, and metrics.
• GenAI semantic normalization for model, provider, tool, workflow, token, cost, and error fields.
• Trace rollups, conversation summaries, user summaries, waterfalls, span detail, and attachments.
• DuckDB-backed local telemetry and Cloudflare R2 Data Catalog / Iceberg-backed Worker telemetry

with retention support.

Use Observability when you need:

• To inspect what happened inside a multi-step agent run.
• To compare providers, models, tokens, costs, latency, and error categories.
• To trigger workflows from matched trace conditions instead of manually scanning logs.

Eval Workflows

Workflows turn production evidence into repeatable review loops.

Eval workflow capabilities include:

• Schedule, trace-match, manual, and Gateway usage event triggers.
• Rule-based and LLM-as-judge evaluators against trace or span evidence.
• Standardized evaluation result logs and score artifacts.
• Dataset example extraction from failure traces.
• Slack, Teams, Lark, and email notifications.
• Finding and dataset writes to Notion, Lark Base, monday.com, or Airtable.

Use Eval Workflows when you need:

• Production failures to become review artifacts.
• Regression examples grounded in real traces.
• Lightweight quality gates before prompt, model, or routing changes are accepted.

Cost Governance

Provon is designed to control spend before the model call leaves your system.

Cost governance capabilities include:

• Request, token, and cost policies by organization, project, API key, and model scope.
• Preflight cost/token estimation and reservation.
• 429 blocking on budget or rate-limit violation.
• Actual usage settlement after successful responses.
• Usage threshold and violation events that can trigger workflows.

Use Cost Governance when you need:

• Budgets to be enforced on the hot path, not discovered at invoice time.
• Cost-aware or budget-aware routing across multiple providers.
• Workflow alerts when usage approaches or crosses a policy.

Guardrails

Guardrails run close to the Gateway path.

Guardrail capabilities include:

• PII and secret-leak checks on input and output.
• Estimated input token limits.
• Message history size limits.
• Max completion token limits.
• Repeated tool call, repeated tool result, and tool error loop detection.
• Monitor, block, degrade, abort-stream, and stop-retry style actions.

Use Guardrails when you need:

• Common silent-failure patterns to become visible.
• Tool-loop and history-growth issues to be stopped before they burn time or budget.
• Sensitive payload patterns to be detected before or after model calls.

Workflow Recipes

These recipes chain the capabilities above into end-to-end production loops:

• Match expensive traces, run an evaluator, and write failed cases to Notion.
• Trigger on Gateway budget violations and notify the owning team.
• Promote a failure trace into a dataset example for regression testing.
• Compare prompt/model variants in Playground before accepting a change.
• Route simple workloads to cheaper targets and preserve expensive models for harder tasks.

Related Docs

• Model providers
• Telemetry
• Deployment
• Integrations