AI agents you can trust in production.

1. 01

   Six identical prompts

   search_kb found no articles, and resolver-agent sent gpt-5 the same prompt six times. The SDK records the step and the tool, the gateway every model call.

   • workflow_runsupport-ticket-resolutionone run, two of its steps

     one run, two of its steps

     • agent_stepresearchresolver-agent

       resolver-agent

       • tool_callsearch_kb · emptykb-vector-store · 0 articles

         kb-vector-store · 0 articles

     • agent_stepdraft_replyresolver-agent

       resolver-agent

       • llm_callsame prompt · 1 of 6chat.completions gpt-5

         chat.completions gpt-5

       • llm_callsame prompt · 2 of 6chat.completions gpt-5

         chat.completions gpt-5

       • llm_callsame prompt · 3 of 6chat.completions gpt-5

         chat.completions gpt-5

       • llm_callsame prompt · 4 of 6chat.completions gpt-5

         chat.completions gpt-5

       • llm_callsame prompt · 5 of 6chat.completions gpt-5

         chat.completions gpt-5

       • llm_callsame prompt · 6 of 6chat.completions gpt-5

         chat.completions gpt-5

2. 02

   Flagged on the fourth send

   agent_loop fires when one run sends the same prompt four times. In our benchmark the incident opened 1.9 s after the looping call (measured — how we measured time to detect).

   detector

   agent_loop

   rule

   same prompt 4× in one run

   seen

   6× the same prompt hash

   severity

   high, critical in production

3. 03

   What the incident carries

   The incident arrives with its evidence: the repeated call, the identical prompt and the empty search that started the loop.

   ev-1

   repeated gpt-5 call

   ev-2

   identical prompt repeated 6×

   ev-3

   search_kb returned no articles

   finding

   step count 2× a healthy run

4. 04

   The next call gets a 423

   Worksi suggests stopping the run or pausing the agent. Once it is paused, its next call is refused at the gateway, before your provider sees it.
   Response

   HTTP/1.1 423 Locked{  "error": {    "message": "agent \"resolver-agent\" is paused",    "type": "invalid_request_error",    "code": "agent_paused",    "param": null,    "aiw_request_id": "req_…"  }}

   Example response. The body format is the gateway’s.

SDK packages are distributed by Worksi. Request access. Any OpenAI or Anthropic SDK works with the gateway without them.

import osfrom openai import APIStatusError, OpenAI client = OpenAI(    base_url="https://gateway.worksi.ai/v1",    api_key=os.environ["AIW_API_KEY"],    default_headers={"X-AIW-Agent": "resolver-agent"},) def draft_reply(run_id: str, history: list) -> str | None:    try:        reply = client.chat.completions.create(            model="gpt-5",            messages=history,            # ties every call of the run together            extra_headers={"X-AIW-Run-Id": run_id},        )    except APIStatusError as error:        # 423 agent_paused or 409 run_stopped: hand over        if error.status_code in (409, 423):            return None        raise    return reply.choices[0].message.content

from aiworkforce import AIW, AgentPausedError, RunStoppedError aiw = AIW(service="resolver-agent")client = aiw.openai() def resolve(ticket: Ticket) -> None:    with aiw.run("support-ticket-resolution",                 agent="resolver-agent") as run:        with aiw.tool("search_kb", args={"q": ticket.subject},                      data_source="kb-vector-store") as t:            articles = search_kb(ticket.subject)            t.set_result({"articles": len(articles)})        try:            with aiw.step("draft_reply"):                client.chat.completions.create(                    model="gpt-5",                    messages=prompt(ticket, articles),                )        except (AgentPausedError, RunStoppedError):            # paused or stopped from the console            return hand_over(ticket)        run.outcome("ticket_resolved", success=True)