DBOS vs Kestra (2026)

Native integrations with Pydantic AI, LlamaIndex, the OpenAI Agents SDK, Google ADK and the Vercel AI SDK make the customer's agent framework durable instead of supplying the tools. They wrap the agent loop as a workflow and each model request, tool call and MCP communication as a checkpointed step, and any step can wrap an outside API call with retries. The tools, credentials and connectors belong to the framework and the customer's code, and DBOS has no connector catalog or framework of its own for authenticated actions.

Sourcedocs.dbos.dev/integrations/pydantic-airead 2026-09-23

The AI Agent task takes tools of Kestra's own, including any Kestra task or flow called as a tool (KestraTask, KestraFlow), web search (Google Custom Search, Tavily), MCP clients over SSE, stdio and streamable HTTP, and another agent as a tool, and the platform ships over 2,000 plugins for databases, cloud services, messaging and business systems with managed credentials through secrets.

Sourcekestra.io/docs/ai-tools/ai-agentsread 2026-09-25

Durable orchestration is the product. Workflows and steps are written as ordinary code in Python, TypeScript, Java or Go, with automatic step retries, durable queues carrying concurrency and rate limits, child workflows, durable sleeps and waits, messages and events between workflows, and agent loops (including subagents through the Vercel AI SDK integration) run as workflows.

Sourcedocs.dbos.dev/ai/ai-quickstartread 2026-09-23

Kestra is a workflow engine whose flows, declared in YAML, run tasks through flowable control primitives (Sequential, Parallel, Loop, LoopUntil, Switch, If, DAG and Subflow), with retries, error handling and backfills. For agents, an AI Agent task decides which tools and tasks to run and in what order, a parent agent can delegate to a child agent wrapped as a tool, and an A2A client forwards work to a remote agent.

Sourcekestra.io/docs/workflow-components/tasks/flowable-tasksread 2026-09-25

Work starts on its own. Scheduled workflows run exactly once per interval on cron schedules, which can be created at run time and are stored in the database, and durable queues start enqueued workflows as capacity allows.

Sourcedocs.dbos.dev/golang/tutorials/scheduled-workflowsread 2026-09-23

Flows declare Schedule (cron), Flow (another flow finishing), Webhook, Polling and Realtime triggers, and an MCP Tool trigger registers a flow as a tool an outside agent can call.

Sourcekestra.io/docs/workflow-components/triggersread 2026-09-25

There is no retrieval structure DBOS maintains over the customer's corpus. Workflows are checkpointed to the customer's own Postgres, and the Vercel AI SDK integration makes embedding calls durable steps, but the index and retrieval belong to the customer's code.

Sourcedocs.dbos.dev/examplesread 2026-09-23

The IngestDocument task chunks and embeds the customer's documents into an embeddings store (Kestra's own KV Store, or Chroma, Elasticsearch, PGVector, Pinecone, Qdrant, Weaviate and others), and the RAG ChatCompletion task and EmbeddingStoreRetriever query that store so an agent answers from the customer's data. The embeddings persist across runs.

Sourcekestra.io/docs/ai-tools/ai-rag-workflowsread 2026-09-25

State is kept inside one run. Every workflow and step is checkpointed to Postgres, so a workflow that crashes resumes from its last completed step and the agent's code carries on from the restored outputs. The scope is one run, so a checkpoint lets the same run finish and carries nothing into the next. Step outputs sit in queryable tables, and Conductor shows the history. There is no state store that persists across runs for each user, conversation or entity as agent memory.

Sourcedocs.dbos.dev/production/retentionread 2026-09-23

Agents built in Kestra remember information across executions through a memory plugin. The KestraKVStore memory saves the chat messages from the agent's own turns to a namespace KV entry keyed by memoryId, with PostgreSQL and Redis memories as alternatives, so one conversation or customer can be deleted without touching the rest. A TTL sets its expiry, drop set to AFTER_TASKRUN deletes it after use, and scheduled purges cover KV pairs. KV pairs can be created, edited and deleted in the Kestra UI, through tasks and through the REST API.

Sourcekestra.io/plugins/plugin-ai/memory/io.kestra.plugin.ai.memory.kestrakvstoreread 2026-09-25

Human input is handled with a durable wait in the customer's code. In the human in the loop guide, an agent calls DBOS.recv to wait hours or days for a notification with a deadline, checkpointed so the wait survives restarts and upgrades, while DBOS.send delivers the person's response and workflow events publish the agent's status. That is a mechanism the customer's code calls, but it is general durable messaging applied to approval, with no approval step, approver identity, review surface or policy guardrail of DBOS's own, and the gate's logic is the customer's code.

Sourcedocs.dbos.dev/ai/hitlread 2026-09-23

A Pause task holds a flow at a chosen step, for output validation or manual approval, until a person resumes it from the UI or API. Approval chains route decisions to specific users or teams, with audit logs recording who approved or rejected each request and why, and Human-in-the-Loop Approvals with custom inputs come in the Enterprise Edition.

Sourcekestra.io/docs/use-cases/approval-processesread 2026-09-25

DBOS states SOC 2 compliance, audited regularly, along with GDPR and CCPA, with a HIPAA BAA available and SOC 2 and HIPAA support on Teams. Teams adds role based access control and Enterprise adds SSO and SAML, and Conductor's organization management supports RBAC and any OAuth compatible SSO. DBOS also states that Conductor is off the execution path and never sees workflow data.

Sourcedbos.dev/dbos-pricingread 2026-09-23

The Enterprise Edition includes SSO, LDAP and SCIM, role based access control with a permissions reference, policies, and multi tenant storage isolation, with audit logs and a secrets manager as supporting controls. Kestra Cloud is SOC 2 compliant.

Sourcekestra.io/pricingread 2026-09-25

Each workflow is recorded step by step. Every step's input and output is checkpointed, and the Conductor console shows workflow state, history and traces, monitors queues in real time and alerts on failures (custom alerting on Teams), with managed retention policies setting how much history each application keeps and for how long. The library emits OpenTelemetry traces and logs, OpenMetrics is supported on Teams, and an MCP server lets the customer's own agents inspect workflows.

Sourcedocs.dbos.dev/production/conductorread 2026-09-23

Every execution is recorded for step by step inspection. The Executions view shows status, logs, outputs and metrics per task through Gantt and Topology views, AI Agent tasks can log each model request and response, and the Enterprise Edition adds audit logs recording every action by users and service accounts plus log shipping to an external store.

Sourcekestra.io/docs/ui/executionsread 2026-09-25

The customer decides where DBOS runs. The open source library runs inside the customer's application on any Linux, macOS or Windows host, on premises or in any cloud, against the customer's own Postgres, and Conductor can be self hosted (with Kubernetes guides) and run in air gapped environments on Enterprise.

Sourcedocs.dbos.dev/production/hosting-conductorread 2026-09-23

The open source edition is Apache 2.0 licensed and self hosts on the customer's infrastructure, cloud or Kubernetes, and the Enterprise Edition runs on cloud, on premises or air gapped, with Kestra Cloud as the managed option.

Sourcekestra.io/pricingread 2026-09-25

Quickstarts and an examples library ship in each language, including AI examples such as an agent inbox and a customer service agent. These are examples a developer builds from, not a catalog of agents a customer adopts.

Sourcedocs.dbos.dev/examplesread 2026-09-23

Kestra publishes Blueprints, a searchable catalog of validated, documented flows, including AI agent and RAG flows, that a buyer adopts into a namespace with one click from inside the product and then adjusts; the Enterprise Edition adds custom blueprints, an organization's own approved set.

Sourcekestra.io/docs/concepts/blueprintsread 2026-09-25

No model runs inside DBOS for a customer to choose. The model lives in the customer's own framework (Pydantic AI, the OpenAI Agents SDK, Google ADK, the Vercel AI SDK), and DBOS checkpoints each model request as a step whatever the model is. Being model agnostic because the model belongs to someone else is different from offering a choice of model.

Sourcedocs.dbos.dev/integrations/vercel-airead 2026-09-23

Each AI task takes its model through Kestra's provider property, so the customer picks among providers of Kestra's own for Anthropic, OpenAI, Azure OpenAI, Amazon Bedrock, Google Gemini and Vertex AI, Mistral, DeepSeek, Ollama, OpenRouter and others, and an Enterprise policy can set the provider across a namespace.

Sourcekestra.io/docs/ai-tools/ai-agentsread 2026-09-25

Outside systems can call DBOS through open source SDKs for Python, TypeScript, Java and Go, a Conductor API with a published OpenAPI specification for managing workflows in code, the dbosctl command line tool, and an MCP server for inspecting workflows.

Sourcedocs.dbos.dev/production/conductor-apiread 2026-09-23

Kestra is API first, with every UI action available through its REST API, and ships official SDKs for Java, Python, JavaScript and TypeScript, and Go, a Terraform provider, and custom plugin development. Its MCP servers expose flows as named tools over HTTP for outside agents, with authentication set per server.

Sourcekestra.io/docs/api-reference/kestra-sdkread 2026-09-25

Debugging after a failure is supported. The debugging guide uses the checkpointed history to find where an agent went wrong and workflow fork to restart it from a chosen completed step, reproducing the state up to that point deterministically, in code or from the Conductor UI. That is debugging, not testing. There is no evaluation harness, scored test cases or quality gate of DBOS's own, and the example evaluation step is the customer's own LLM call.

Sourcedocs.dbos.dev/ai/debuggingread 2026-09-23

Unit Tests in the Enterprise Edition run test suites that mock tasks and assert on flow behavior, so a change can be checked before production without real side effects. Those assertions check that the flow behaves as configured, and Kestra does not describe an evaluation harness, scored test cases or a judge for agent output.

Sourcekestra.io/docs/enterprise/governance/unit-testsread 2026-09-25

DBOS operates no browser or desktop for the agent and has no browser automation environment of its own. It makes the customer's code durable wherever it runs.

Sourcedocs.dbos.dev/ai/ai-quickstartread 2026-09-23

The AI Agent's tools are Kestra tasks and flows, web search, MCP clients and other agents, none of them a browser or desktop it operates. Kestra does not describe a browser automation environment or a hosted browser workload, though script tasks run any code the customer supplies.

Sourcekestra.io/docs/ai-tools/ai-agentsread 2026-09-25