A practitioner guide from Alibaba Cloud — documentation, not a model, paper or framework. It is an open-source white paper on building, running, governing and tuning enterprise AI agents, written by Alibaba Cloud staff and maintained as a GitHub repository under Apache 2.0. It succeeds the company's September 2025 AI-Native Application Architecture White Paper, restructured around three problems the authors say have replaced "build an agent quickly": turning probabilistic intelligence into reliable productivity, scaling from isolated experiments to shared infrastructure (stability, security, performance, cost), and moving agents into core business processes. Thirty chapters sit in seven parts — Architecture; Building (harness construction patterns and responsibility boundaries, task orchestration and long-horizon continuation, context/state/reusable assets, controlled execution with Function Calling, MCP and A2A); Runtime (sandboxes, event-log and checkpoint state storage, AI gateways, asynchronous tasks, multi-agent coordination, distributed communication); Governance (observability, prompt-injection and authorization security, AI-asset registration and versioning, pre-release simulation); Optimization (model tuning, trajectory data, golden datasets, bad-case loops, controlled self-evolution, edge runtime); Practice (16 named enterprise case studies, including Geely, Bilibili and MiniMax); and a closing chapter on "Agentic OS". The chapters are in Chinese; the English README is a navigation guide and states explicitly that it is not a translation.

The repository root also carries a 2026 Agent Developer Survey Report based on 1,906 valid questionnaires collected at in-person developer events in Beijing, Shenzhen, Guangzhou, Shanghai and Hangzhou, with respondents described as mostly enterprise technical decision-makers, architects, engineers and product managers. Reported findings include: 46% have an agent built or in development against 36% a year earlier, and respondents with no plans fell from 22% to 15%, but only 18% have deployed one to production; 40% run a single agent as the primary architecture, 42% are adopting multi-agent as primary and 15% design human-in-the-loop as the primary pattern; 57% buy coding agents (Claude Code, Codex, Cursor, Qoder are the examples given) and 78% use a general agent framework; 90% report explicit context- and memory-management needs, and 63% name multi-model routing with cost governance as the capability they most need to add — the single highest item in the survey; 55% still judge agent quality by manual spot-checks and under 8% can evaluate automatically from execution trajectories; and about one in ten reach a task success rate above 90%, 55% above 70%, while 16% have never measured it at all.

Apache 2.0, about 1,100 GitHub stars and 160 forks. There is no code, model, dataset or benchmark in the repository — it is prose, figures and case studies, tracked here for the survey data and for the harness/context/governance taxonomy it puts on the record.

Library

License Apache 2.0
agentsagent-harnessagenticenterpriseinfrastructureopen-source