2pm.spaceAgents that do the work,
not just answer a prompt
Start with one prompt, or draw the job as a workflow: find the right documents, reason and call tools, check the answer, ask a person when it matters, and reply. Every run shows what each step did, what it spent and why.
From idea to working agent in three steps
Draw it, equip it, put it where the work is.
Draw the job
Pick Simple for a single prompt, or Agentic to lay the job out as nodes on a canvas — retrieval, reasoning, checks, branches — wired left to right.
Give it what it needs
Attach tools, skills, knowledge and memory to the agent or to a single node, and set the guardrails every message passes through.
Put it to work
Test it on the canvas, then connect it to an inbox channel, a schedule, another agent as a tool, or an MCP client.
Draw the job
Pick Simple for a single prompt, or Agentic to lay the job out as nodes on a canvas — retrieval, reasoning, checks, branches — wired left to right.
Give it what it needs
Attach tools, skills, knowledge and memory to the agent or to a single node, and set the guardrails every message passes through.
Put it to work
Test it on the canvas, then connect it to an inbox channel, a schedule, another agent as a tool, or an MCP client.
More than a prompt with tools
A canvas for the jobs one prompt cannot do on its own.
Simple or agentic
A simple agent is one model call with your system prompt — right for questions, translation and summaries. An agentic agent is a graph of nodes, for jobs that take several steps and a decision between them.
Sixteen kinds of step
Agent, Crew, Run Agent, Code, Knowledge and Skill Retrieval, Drive Action, Condition, Parallel, Loop, Evaluation, Human Review, Message, Send — each a card you drag onto the canvas and wire to the next.
It checks its own work
An Evaluation node scores the answer with rules or an LLM judge and routes it: Pass moves on, Retry sends it back to the agent for another attempt, up to the retries you allow.
A person where it matters
A Human Review node pauses the run until someone approves or rejects it, and a Message node can ask the customer for details through a form or buttons before the flow goes on.
Tools, skills and knowledge
Sixty built-in tools, your own HTTP APIs and Python, MCP servers, and other agents as tools. A skill loads into every prompt, or only when a question matches it — so a large library does not cost tokens on every call.
Browse the tool catalogTest it, trace it, roll it back
Run the flow from the canvas and read each node's output, tokens and cost. Save canvas versions, and restore one when a change does not hold up.
What's on the canvas
Every node, the model catalogue, and every place an agent can run.
Reasoning & actions
- Agent — its own model, prompt and tools, in a reason-and-act loop
- Crew — a CrewAI crew of agents and tasks
- Run Agent — call an agent you already built
- Code — Python or Bash in a sandbox, no LLM cost
Flow control
- Start — where a message enters
- Condition — branch on a rule, no LLM cost
- Parallel — every outgoing branch at once
- Loop and Exit Loop — over a list or N times
Checks & people
- Evaluation — rules or an LLM judge, then Pass or Retry
- Human Review — pause for approve or reject
- Message — a chat message, a form or buttons
- Send — message the customer and carry on
Knowledge & data
- Knowledge Retrieval — semantic, keyword or hybrid search
- Skill Retrieval — match skills without an LLM step
- Drive Action — create, read or update docs, tables, mind maps and storyboards
- Memory that lasts between conversations, and a workspace memory every agent shares
Models
- Anthropic, OpenAI, Google and DeepSeek
- Alibaba Qwen, Z.AI GLM, xAI and MiniMax
- A model per Agent node, not one per workflow
- Billed from workspace credits — no provider keys to manage
Where it runs
- Inbox channels: Messenger, Telegram, personal Zalo, website chat
- The Scheduler, on a recurring timetable
- Another agent, exposed as a tool
- MCP clients such as Claude Code and Cursor
- The Playground, before any customer sees it
A prompt box vs. an agent builder
The same job, built two ways.
One prompt tries to retrieve, reason, check and reply at once — and you can't see which part failed.
Each job is its own node, and the run log shows what every node received, returned and spent.
A wrong answer goes straight to the customer.
An Evaluation node catches it and sends it back for another try before anyone sees it.
A risky action needs a developer to build an approval step.
Put a Human Review node in front of it, and the run waits for a yes.
A change that breaks the agent means rebuilding it from memory.
Restore the canvas version from before the change.
Agent Builder questions
What people ask before they build their first workflow.
Build the agent your work actually needs
Start with a prompt. Grow it into a workflow when the job asks for one.