Skip to main content

Credit-based plans and reliability fixes

Credit-based plans​

Usage on Agenta Cloud is now counted in credits. The plan cards, the billing page and the Plans and credits page show what each plan includes.

  • Starter is the former Pro plan. It keeps the same price and the same 2,900 credits a month, so existing Pro customers see a new name and no other change.
  • Business now includes 32,000 credits a month.
  • Daily free credits (75 a day, on up to 10 days a month) now go to Hobby, Starter and Business organizations. You can see them as "Daily free credits" in Settings > Credits.
  • New organizations start on Hobby with no trial, and receive a one-time 500-credit "Signup bonus". Existing organizations do not receive the bonus.
  • Concurrent tasks are 2 on Hobby, 20 on Starter and 50 on Business. A Hobby organization can have one project.

When an organization has no credits left, a new task does not start and is not charged. The chat shows "You're out of credits" and tells you the next step for your plan. Tasks that are already running finish.

Fixes​

  • Analytics: a 30-day range that takes too long now shows an error with a message to choose a shorter range. Before, the charts stayed empty. Agent runs also keep their model and connection mode on analytics, so the model charts and the model filter include new runs again.
  • Live replies: in /m, a reply in a session that belongs to a project other than your default project now streams while the model writes, instead of appearing at the end.
  • Gemini on Vertex AI: an agent that uses the pi_core harness can now run Gemini models from a Vertex AI connection, including runs with several tool calls in a row.
  • Triggers: schedules and event triggers that you add from an agent's config panel now bind to the agent by its slug.

For self-hosters​

Upgrade notes for version 0.122.5. The configuration reference lists every variable named here.

  • Span and trace queries return one page. POST /spans/query, POST /traces/query and the legacy POST /tracing/spans/query now return 100 rows when the request sets no windowing.limit, and at most 1000 rows. When more rows match, the response has a windowing object. Send it back as windowing to read the next page. A client that expected every row in one response must follow this cursor. You can change the limits with AGENTA_API_TRACING_QUERY_DEFAULT_LIMIT and AGENTA_API_TRACING_QUERY_MAX_LIMIT. See the query pattern guide.
  • Custom workflow URLs are checked. The API now applies the gateway outbound rules to the URL of a custom workflow or evaluator, and service calls no longer follow redirects. A custom workflow on a private network or on plain HTTP needs AGENTA_GATEWAYS_INSECURE_EGRESS_ALLOWED=true, or its host in AGENTA_MCP_GATEWAY_HOST_ALLOWLIST. The Docker Compose env templates already set the flag. The Helm chart and the Railway templates leave it unset. A custom workflow that answers with a redirect now fails, so point its URL at the final address.
  • Longer gunicorn keep-alive. The Helm chart and the Docker Compose gh stacks now run gunicorn with --keep-alive 620, which is longer than the idle timeout of the common load balancers. This removes occasional 502 errors on a reused idle connection. In the Helm chart you can set it with api.gunicorn.keepAlive and services.gunicorn.keepAlive.
  • Smaller, configurable Postgres pools. Each process now sizes its pools from POSTGRES_POOL_SIZE, POSTGRES_MAX_OVERFLOW, POSTGRES_WORKER_POOL_SIZE and POSTGRES_WORKER_MAX_OVERFLOW. At peak, an api process uses up to 30 connections and a worker process uses up to 20. Set max_connections on your database for all of your processes together.
  • Daytona sandbox snapshot. If you use Daytona, build the snapshot agenta-agent-sandbox-v3 before you upgrade the runner. Until the snapshot exists, new sandboxes fail to start. If you pin AGENTA_RUNNER_DAYTONA_SNAPSHOT, do not pin it to agenta-agent-sandbox-v2.