Plan
Agent breaks the task down: goal, constraints, available tools, relevant knowledge. Humans can edit the plan before execution.
- Explicit plan steps
- Edit plan in human
- Replanning supported
Product · AiDo
Turn LLM capability into
executable, traceable, collaborative digital workers.
AiDo is the enterprise agent suite on AI-OS. Plan / Action / Reflection makes every agent output traceable; durable runs and session replay make long tasks and post-mortem reviews everyday workbench work.
PAR · Three-phase agent
An agent is not 'one prompt, one answer'.
Break-down → execute → reflect — AiDo makes these three things mandatory flow. Customers see what the agent thinks, what it does, and how it judges itself afterward.
Action
Calls tools, reads knowledge, writes documents, sends emails, updates tickets according to the plan. Every step has structured I/O records.
- Whitelist-bound tool calls
- Structured call logs
- Human approval nodes
Reflection
After execution agent self-checks: was the plan met? Did output meet standards? Error signals flow into evaluation for next run.
- Auto eval scoring
- Exception rollback
- Eval data into prompts / knowledge
Runtime · Built for production
Long-running, resumable, collaborative runtime.
Enterprise agents can't end in two minutes. AiDo is designed around long tasks, pause / resume, replay and parallel sessions so agents become a daily workbench.
Durable runs
Browser refresh, tab close, network drop don't kill long tasks. Events persist to disk; resume from the same point next time.
Session replay
Every session is archived as rollout JSONL. Any task can be replayed at original speed, annotated, edited and re-run.
Pause / interrupt
Customers can pause, ask, or change direction at any time — agent holds an internal abort controller, no runaway behavior.
Parallel sessions
One user, many agents in parallel. Sessions don't fight for resources or leak data across each other.
Lifecycle
From natural language to archived session.
Every step structured. Audit, replay and improvement are all evidence-based.
Task received
User asks in natural language with knowledge references and goal.
natural language
Plan emerges
Agent emits plan first; user can confirm / edit / replan.
explicit plan
Action executes
Calls tools, reads knowledge, modifies documents, triggers external APIs. Every step recorded.
tool calls
Reflection
Agent self-checks completion; user marks satisfaction; eval signal flows back.
self-eval
Archive & replay
Session lands in archive — replay, annotate, edit, refine prompt anytime.
rollout JSONL
Tools · Integrations
Connects to what you already run.
AiDo isn't an island agent — it uses your existing email, docs, tickets, code, APIs. Every call passes the whitelist and audit chain.
Email / calendar / tickets
Connects to enterprise mailbox, calendar, Jira / DingTalk / Feishu tickets.
Docs / knowledge
Confluence / Notion / Feishu docs / self-hosted wiki / knowledge space.
Code / DevOps
GitHub / GitLab / Jenkins / Argo / any SaaS API.
Private APIs
REST / GraphQL / gRPC inside your perimeter, authorized via tool whitelist.
On-Chain Identity
Want regulator-verifiable agent decisions? Hook into Shuke Chain.
AiDo agents can run inside ordinary business — or carry blockchain identities and become regulator-auditable digital workers.
Through Shuke Chain's Agent on Chain, an AiDo agent registers an independent account on a FISCO BCOS SM-cryptography consortium chain (private key vaulted by OpenBao). Every contract call it makes is SM2-signed and on-chain. For tourism, supply chain, IP rights, carbon assets — any scenario that needs AI auto-decision + regulator traceability — this upgrades Plan / Action / Reflection from internal logs to an immutable ledger.
FAQ
Studio relationship / long tasks / mishandling / models — five common questions.
How does AiDo relate to AiDo Studio?
AiDo is the agent suite — ready-made enterprise agents (task assistant, doc assistant, ops assistant). AiDo Studio is the agent factory where employees and teams build agents. They share data: Studio creations can be one-click published into AiDo.
Does AiDo require AI-OS?
Yes. AiDo strongly relies on AI-OS for identity, knowledge space, model gateway, tool whitelists and audit. These are the prerequisites for governable agents — a stand-alone agent framework can't deliver them.
What happens if a long task is running and the client disconnects?
AiDo's durable runs handle this: browser refresh, tab close or network drop don't interrupt the backend runner. Events persist to disk and you resume seamlessly. See AI-OS docs/toHeroTask.md §11 for the protocol.
Can agents mishandle tools (delete / send by mistake)?
Three lines of defense: ① tool whitelist by agent type; ② high-risk operations (delete / send / transfer) go through human approval; ③ full audit log with one-click takedown by admins. Code agents must run in a sandbox.
Which models? Will swapping models affect sessions?
Models flow through the AI-OS gateway. Today Instant / Pro are public; enterprises can mount private models. Model swaps only affect future turns — existing sessions stay replayable in their original form.
Try AiDo
Spend 30 minutes with an AiDo agent.
aios.shukeyun.com/demo offers a no-login experience. A plan → action → reflection run takes 15-30 minutes. Want to build your own? Head to AiDo Studio.