builders · review · first seen Nov 6, 10:33 UTC

CrewAI adds experimental conversational Agent sessions

CrewAI 1.15.36 shipped experimental conversational Agent sessions and changed how crew runs handle a missing API key. The reason to review it today is the change to a run’s failure path; the session feature still needs evaluation before a builder depends on it.

L0 · primary1 source4 claims checked4 min read

What shipped

The release adds experimental conversational Agent sessions. That is the full feature description in the supplied release excerpt. It does not describe a session API, configuration, persistence model, or example interaction. I would inspect the addition in a test project, rather than assume an existing crew can use it without changes. Experimental is the release’s label, not a measured assessment of stability. [1]

The release changes two paths involving a missing API key. When crewai run starts, it asks for the missing key. A JSON crew run with a missing key now stops cleanly. I read these as separate changes: a request for the key at startup and a clean stop for the JSON run. The excerpt also mentions a fix to empty-value precedence in the run’s key check, but gives no reproduction steps or description of which value takes precedence. [1][2]

Tool calls for GPT-6 Astra and GPT-6.1 Sol now go through the Responses API, according to the release. This is a specific routing fix, not evidence that every model or every operation uses that API. The excerpt provides no request trace or compatibility matrix. If either named model is in your tool-calling path, include that path in your checks. Otherwise, this item alone gives you no stated reason to change your project today. [1]

Why it matters if you build

For a builder starting a crew with crewai run, the missing-key request changes what happens at startup. For a JSON crew run, the release says the missing-key case ends cleanly. I would test those paths separately rather than treat one result as proof of the other. The notes do not say whether the request works in unattended execution, how a supplied key is retained, or what the clean stop prints or returns. Do not build automation around an assumed exit code or prompt format. [1][2]

Conversational sessions are worth examining if an ongoing Agent interaction is relevant to your workflow, but the release establishes only that an experimental feature was added. It does not establish how continuation, state storage, concurrency, or recovery work. That supports a contained evaluation, not a production dependency. For the named model tool calls, the concrete change is the Responses API route. The notes give no cost, latency, or output-quality comparison, so those tradeoffs remain questions for your own tests. [1][2]

How to try it today

1. Inspect the 1.15.36 release entry for the experimental conversational Agent sessions item.
2. In a test project using 1.15.36, examine session behavior before depending on it.
3. Start crewai run without its API key and observe the request for the key.
4. Run a JSON crew without its API key and observe whether it stops cleanly.
5. If you use GPT-6 Astra or GPT-6.1 Sol tool calls, test that path and check its Responses API routing.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. shipped L0 · primary

   CrewAI 1.15.36 adds experimental conversational Agent sessions.

   per GitHub releases · crewAIInc/crewAI (L0)
2. shipped L0 · primary

   CrewAI 1.15.36 asks for a missing API key when crewai run starts.

   per GitHub releases · crewAIInc/crewAI (L0)
3. shipped L0 · primary

   CrewAI 1.15.36 stops a JSON crew run cleanly when its API key is missing.

   per GitHub releases · crewAIInc/crewAI (L0)
4. shipped L0 · primary

   CrewAI 1.15.36 calls GPT-6 Astra and GPT-6.1 Sol tools through the Responses API.

   per GitHub releases · crewAIInc/crewAI (L0)

What is not confirmed

The release does not specify how to create or resume a conversational Agent session, how long one lasts, where it stores state, or what experimental means for compatibility. It does not specify the missing-key request’s behavior without an operator, the JSON run’s exit code, or key-handling security properties. It gives no measured cost, latency, reliability, or quality result for the Responses API routing. Those are open questions, not release promises. [1][2][3][4]

Who should act, who should wait

Move now to test 1.15.36 if a missing API key is disrupting your runs or tool calls to either named model are already in your workflow. Reproduce the exact path you rely on and check its observed behavior. Evaluate conversational sessions in isolation if they address a near-term need. Wait before making that experimental feature a required production dependency, or assuming the new key request solves unattended runs: the supplied notes establish neither property. [1][2][3][4]

Sources

• 1.15.36 — GitHub releases · crewAIInc/crewAI L0 · primary