AI coding changed in the last year. Describe a feature to Claude Code or Codex and it designs, implements, tests, checks its own work and reports back.
Yet when people build AI agents or workflows, the default is still a drag-and-drop flow builder such as Dify. That is great for a prototype. For an agent that runs a real business process, it starts to hurt:
- permissions (who may read what), approvals (who decides) and exceptions are hard to express as boxes and arrows;
- reviewing, testing and rolling back a flow change — routine in software — is awkward;
- the most capable developer we now have, the AI coding tool, can barely touch a flow that lives in a canvas.
Biz AI Agent takes the opposite approach. It is an open-source AI agent framework (Apache-2.0):
An agent is first an ordinary IT system — so you build it with AI coding.
- Repository: https://github.com/ITC-cloud-soft/biz-ai-agent
- Website: https://biz-ai.myaiportal.net/en/biz-ai-agent.html
1. A skill plus an AI coding tool
The repository ships the agent-builder skill for AI coding tools. Clone it and open your tool at the root — the skill is picked up automatically.
git clone https://github.com/ITC-cloud-soft/biz-ai-agent.git cd biz-ai-agent && npm install claude # Claude Code loads .claude/skills/agent-builder # codex # Codex discovers .agents/skills/agent-builder Then ask in your own words:
Use the agent-builder skill to build a "travel expenses" AI agent: - employees ask about the expense rules and check their own claims on Slack; - managers approve their team's claims; above $1,000 finance approves too; - finance gets a monthly count report; - the rules are in docs/travel-rules.md. Following the skill, the AI:
- asks people first — who uses the agent, what it may read, who approves, who may read which document, what a correct answer is — writes it into
AGENT_PLAN.mdand waits (no guessing); - writes the business system and its manifest and passes the Bridge v1 conformance suite: 52 / 52;
- adds permission-aware knowledge, the "why did this reach me?" graph, evals written before any tuning, the settings screens, SNS approvals and the DeepSeek Harness (DSH) profile;
- verifies every step with commands and reports the results, what is not verified and what is still to confirm.
examples/equipment-loans in the repo was built from scratch by an AI that followed only this skill.
A no-code tool is for people drawing flows. This is AI writing code, with tests guarding it. Branches, permissions and exceptions are plain code and tests — there is no ceiling on what the agent can do.
2. Build, deploy and maintain it like a web app
What the AI produces is an ordinary Node.js project in examples/<system-id>/:
| file | what |
|---|---|
bridge.manifest.yaml | the system's self-description: roles, events, the AI's tools |
src/app.mjs | the business system: data, sign-in, business rules, API, settings screens, knowledge, graph |
knowledge/*.md | company documents, with their visibility written inside |
evals/, test/ | evals and tests for business rules, permissions and conformance |
dsh-profile/ | the AI plane's configuration |
Deploying is the same as any website: NODE_ENV=production node src/server.mjs behind an HTTPS reverse proxy. The AI plane (DSH) is one more process on the same machine, reached only through a gateway inside the web app.
Maintenance is the same too: git, reviews, tests and CI just work. Want a change? Ask the AI coding tool again — the tests (52 conformance checks, business rules, permissions, eval floors) tell you nothing broke.
3. Talk to it on chat — and it talks to you
The AI employee works in the chat your staff already use: LINE, Slack, Teams, WeCom, DingTalk, Feishu and more (13 SNS).
Employee → agent
- "What is the daily allowance for a business trip?" — it searches company documents with that employee's own rights and answers with sources (links to the business screens). What you cannot see, the AI cannot see.
- "Why did this approval reach me?" — it explains, from the manifest.
Agent → employee
- a claim comes in → an approval card goes to the approver only (never the requester); one tap to approve;
- business changes are notified only to the people the business system names;
- reports with counts only, on schedule.
It answers when asked and reaches out when needed — two-way, proactive agents. People still decide: the AI's only write is a proposal; confirming, approving and amounts stay with people, enforced by code and configuration, not by prompts.
4. Fully open source — let's build it together
Apache-2.0: source, manual, skill, examples and tests are all public.
npm test # about 900 tests, 0 fail npm run e2e # a real DSH + mock model + mock LINE: 15 / 15 PASS "I built an agent for this process", "I need it on that chat platform", "this part of the skill is unclear" — issues, PRs and stars are all welcome. People who know the business plus people who drive AI coding tools make smarter agents.
- GitHub: https://github.com/ITC-cloud-soft/biz-ai-agent
- Website (the most complete explanation): https://biz-ai.myaiportal.net/en/biz-ai-agent.html