AI Agent Workflow Examples: Cited Public Deployments (2026)
Vendor pages are full of invented “Acme Corp deployed agent for invoices” examples. The discipline on this page is different. Every example below cites a real public source: a paper, a documentation page, a press release, or a vendor product page that names a real deployment. Where the source does not name a real customer or does not publish a date, the example is not on this page.
| Canonical shape | Deployment | Source | Date | Process page |
|---|---|---|---|---|
| Escalation | Klarna AI assistant Tier-1 customer service agent handling two-thirds of chats in its first month, with a customer-initiated live-agent option retained. Klarna reversed direction in May 2025 and recruited human agents again on quality grounds. | www.klarna.com | Klarna press, 27 February 2024; reversal reported May 2025 Accessed October 2026 | page |
| Branching gateway | Cisco Webex AI Agent Contact-centre AI agent that triages, classifies, runs automated remediation, and escalates to a human agent on out-of-library cases. | www.webex.com | Cisco product documentation (2024 to 2025) Accessed April 2026 | page |
| Async loop | Cognition Devin Long-running coding agent run as a team member: tasks are assigned, the agent returns a pull request, humans review it. Cognition describes its own use as “tag @ Devin with a request and we get a PR we can review and test out”. The wrapper process is BPMN-deterministic; the agent loop is non-deterministic. | cognition.com | Cognition blog, 27 February 2026 Accessed October 2026 | page |
| Async loop | Anthropic Claude Code Engineering assistant that reads the codebase, edits files, runs commands, and stages commits and opens pull requests for review. Documented call-activity wrapper around the agent loop. | code.claude.com | Claude Code documentation (accessed October 2026) Accessed October 2026 | page |
| Linear handoff | Clay (Anthropic customer story) Claygent, Clay's AI research agent, scrapes the public web for a prospect's thought leadership and then drafts a personalised email referencing it. Extraction, enrichment and drafting all sit in the agent lane. No human approval step before send is documented, which is why the send gate on our sales swimlane is labelled as our recommendation rather than as Clay's design. | claude.com | Anthropic customer stories (ongoing) Accessed October 2026 | page |
| Agent-to-agent handoff | OpenAI Swarm handoff primitive Reference implementation of agent-to-agent handoff as a first-class operation. Carried forward into the OpenAI Agents SDK. | github.com | OpenAI Swarm repo (released October 2024; README now directs to the Agents SDK) Accessed October 2026 | page |
| Agent-to-agent handoff | OpenAI Agents SDK handoffs Production successor to Swarm. handoff() abstraction with documented context-passing semantics. | openai.github.io | OpenAI Agents SDK docs (2025) Accessed April 2026 | page |
| Agent-to-human handoff | LangGraph interrupt Human-in-the-loop primitive. The graph pauses on interrupt, the human approves or modifies, the graph resumes. | docs.langchain.com | LangGraph interrupts documentation (accessed October 2026) Accessed October 2026 | page |
| Parallel processing | Anthropic orchestrator-workers (Building Effective Agents) The orchestrator-workers pattern documented in Anthropic's December 2024 paper. Coordinator agent fans tasks out to workers, merges results. | www.anthropic.com | Schluntz, December 2024 Accessed April 2026 | page |
| Branching gateway | CrewAI Process abstraction CrewAI models the work as crews of agents and processes. The sequential process executes tasks in a predefined order; the hierarchical process requires a manager that delegates tasks to agents and reviews their output, rather than tasks being pre-assigned. | docs.crewai.com | CrewAI Processes documentation (accessed October 2026) Accessed October 2026 | page |
| Agent-to-agent handoff | AutoGen group chat (maintenance mode) Microsoft AutoGen's AgentChat API uses a group chat abstraction with named participants. Conversation flow across agents maps onto swimlane semantics. AutoGen is now in maintenance mode: Microsoft states it will receive no new features and directs new users to the Microsoft Agent Framework. The pattern remains instructive; the framework is not the one to start on. | github.com | AutoGen repository (maintenance-mode notice current October 2026) Accessed October 2026 | page |
| Cross-vendor handoff | A2A protocol (Google et al., 2025) Open agent-to-agent protocol over HTTP. Discovery, task delegation, and result return. Cross-vendor handoffs as BPMN message flows across pools. | a2a-protocol.org | A2A protocol announcement, April 2025; Linux Foundation project since June 2025 Accessed April 2026 | page |
| Branching gateway | ServiceNow Now Assist AI service desk: incident triage, knowledge-base search, CMDB-aware enrichment, agent-to-human escalation on out-of-library cases. | www.servicenow.com | ServiceNow product documentation (2024 to 2025) Accessed April 2026 | page |
Where an example has since changed
A cited deployment is a snapshot, not a standing endorsement, and the most-quoted one on this page has moved. Klarna's February 2024 release is the figure everyone reuses: two-thirds of customer service chats, the equivalent work of 700 full-time agents. In May 2025 Klarna began recruiting human agents again, with its chief executive telling Bloomberg that cost had been “a too predominant evaluation factor” and that “what you end up having is lower quality”. By February 2026 the company was running an explicitly hybrid model with human contact as the premium tier. Anyone citing the 2024 number as the current state of that deployment is two years behind it. The customer support intake page carries the full sequence with sources.
Sources for the follow-up: Bloomberg (8 May 2025); Customer Experience Dive (9 May 2025); Customer Experience Dive (20 February 2026).
Inclusion rule
Every entry on this page is a real public artefact: a research paper, a framework documentation page, a vendor product page, a press release, or a customer story page that names a real customer. Invented “Acme Corp deployed AI for X” examples are excluded. Vendor case studies that do not name a customer or do not publish a date are excluded. AI-generated content farms are excluded.
Where a source has moved
A citation is only worth what the link still resolves to, and framework documentation moves faster than the patterns it documents. Checked in October 2026, several sources on this page had relocated and are updated above: the LangGraph documentation now lives on docs.langchain.com (the old langchain-ai.github.io pages are redirect stubs marked noindex), the CrewAI Processes page moved from core-concepts to en/concepts, Anthropic customer stories moved to claude.com, the Claude Code documentation moved to code.claude.com, and Cognition now publishes on cognition.com. The Cognition entry also points at a specific dated post rather than the blog index, because a blog index is not a citation.
Update cadence
Quarterly review of Anthropic, LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, and the major published case studies, including a link-resolution check on every source in the table. As-published addition of new named deployments. Per-example freshness stamp on each entry. See the methodology page for the full review schedule.