Skip to main content
Super Durable
Menu

Sessions

View as Markdown

A durable agent session is one Flow run that holds one conversation. Each model call and each tool call runs in a Step. So a crash, a deploy, or a long wait repeats nothing that already finished.

A durable agent session

One Flow run is one agent session. The Flow ID is the session ID. Dex Agents builds the session Flow: the dex-agents package in Python and @superdurable/dex-agents in TypeScript. A framework adapter adds one agent framework to it.

The session stores the conversation in Attributes. A session waits for the next user message without using a Worker task slot.

Model calls run in a RunModel Step. In most harnesses, each RunModel Step makes one model call, and each tool call runs in its own Step. The Steps of one turn are:

StepWhat it does
RunModelMakes one model call and records the response.
CallToolRuns one call of a read or an idempotent write. Dex may retry it.
CallSingleAttemptToolRuns one call of any other write. Dex runs it at most once.
AwaitToolApprovalWaits for a person to approve or reject a tool call. See Approvals.
ResumeTurnWaits for every tool result of the model response, then moves back to RunModel.

The tool calls of one model response run in parallel. When they all finish, ResumeTurn sends their results to the model.

Latency and cost

Every session Step is SYNC: Dex commits the Step's start before the Worker gets it, and its result before the session moves on, each to object storage. A turn in which the model asks for three tools and then answers takes six Steps: two RunModel, three tool Steps, and one ResumeTurn. Performance measured a p50 of about 105 ms for a SYNC Step on a 2-vCPU server at half of its peak load. The three tool Steps run in parallel, so the turn waits on four Steps in a row. At that p50, they add about 0.4 seconds to the turn, on top of the model and tool time.

A RunModel Step retries up to three attempts within 30 minutes. A retry sends the model call to the provider again, so you pay for its tokens again. run_model_limit in Python, or runModelLimit in TypeScript, lowers the attempts for an expensive call. Step timeouts lists the defaults.

A crash repeats only the Step that was running. Dex does not replay the loop. So the agent code needs no determinism rules, and a code change takes effect at the next Step.

Two harnesses group more work into one Step:

  • Claude Agent SDK. The Claude Code CLI runs a turn segment in one RunModel Step. The segment holds the model calls and the read tool calls, and a failed segment reruns as a whole. Writes leave the segment one at a time, each in its own Step, so they never run in parallel.
  • LangChain and LangGraph. One RunModel Step runs the graph to its next tool calls. A graph can make several model calls on the way, and a retry repeats them together.

What a session remembers

A session keeps one conversation: its messages, its tool calls, and their results. A new session starts empty. Dex Agents has no memory that one session shares with the next.

Keep long-term memory, such as a user's preferences or past orders, in a store that you own. That store can be your database or a Flow that you keep for each user. Give the agent a tool that reads it, and a tool that writes it with the right policy.

A stopped session stays readable for the retention period, and then Dex deletes it. History and retention has the defaults.

Framework support

Framework Support shows how your framework runs a turn, which session features it has, and its limits.

Next