What agents are in GPTree

An agent in GPTree is a persona with a job: a name, a role title, a place in your company, and its own pane you can talk to like a teammate. Setting one up takes about a minute. What it plugs into takes longer to explain, because the agent platform borrows the strongest ideas from the current generation of agent frameworks and puts them inside the one thing those frameworks don't have: a branching workspace where judgment is part of the loop.

A company, not a list of bots

Agents are organized the way people are. The Agents page opens on your org chart ("Your company of agents"), and a template can stand up a whole company at once: the Startup company template creates a CEO with a CMO, CFO, and CTO reporting to her, an engineer and a code reviewer under the CTO, and starter routines already on their calendars.

Structure exists because agents use it. An agent with direct reports has a "Delegate to a teammate" tool that hands work down as a tracked task on the teammate's own queue, with a delegation thread you can open later. Several agents can hold a meeting on one topic and work it in rounds. And crew runs put a group of agents on a single problem, on demand or on their own schedule. If you know CrewAI, the crew idea will feel familiar. Here it's native to the product, and every run lands in the same tree as the rest of your work instead of in a terminal window.

Skills and routines

Every agent has Skills: reusable multi-step pipelines where each step is a plain-language instruction assigned to an agent. Build one in the Skills drawer ("New pipeline"), run it with "Run skill", or let the agent invoke it mid-conversation, because every enabled skill is exposed to the model as a tool it can call. Skill runs can be approval-gated like anything else.

Routines make an agent run without you. Three triggers: a Schedule with a full recurrence builder, a Webhook with a signed URL and a rotatable secret so outside systems can start work, or Manual for run-it-when-I-say-so. Every routine keeps its run history, so "what did it actually do this week" is a click.

Decisions are first-class records

This is the part we think no other agent platform has. Every explicit decision an agent makes is written to the Decisions ledger with its subject, its outcome, and who decided. When two live decisions contradict each other on the same subject, both get flagged and a reconcile item lands in your inbox: you pick the one that stands. When a run isn't confident between paths, the decision trail offers the branch-and-converge move: keep both paths, compare them, then merge the insights into the one you keep. On teams, collaborators can vote on the alternatives before someone resolves the call.

That's the same Branch, Compare, Merge loop the rest of GPTree runs on, applied to agent judgment. GPTree's position is that it supports judgment rather than replacing it, and the ledger is what that means in practice: decisions have provenance, conflicts surface instead of festering, and the record survives the chat that produced it.

Where agents run

In the cloud today. Local execution ships through the bridge integrations: connect a local OpenClaw runtime under Settings and GPTree becomes the control plane for agents running on your own machine, with the loop executing locally and results streaming back into your tree. The product is candid about what's next: the routine dialog shows "Runs on: Cloud" live, with Local sitting beside it marked "soon". We'd rather the interface tell you that than have an article pretend otherwise.

Security is the default posture, not an add-on

Autonomy is off until you turn it on, and it's double-gated: your own autonomy switch plus a global master. High-risk actions (sending email, posting to Slack, payments, anything that executes on a connected machine) stop for your approval every time, with "Approve with edits" for when the action is right but the wording isn't. Each routine carries its own permission level: ask every time, only ask for high-stakes actions, or never ask. Webhook secrets, service-account tokens, and connector credentials are encrypted at rest. Agent memory carries a visible data-residency badge. And every run writes a hash-chained audit timeline, so the record of what an agent did can't be quietly rewritten afterward.

The honest limitation hasn't changed: agents here are not fire-and-forget. Questions, approvals, and conflicting decisions land in the Action Plane, your inbox, and it will pull you back in. That's the design. Our advice for a first agent is still one agent, one narrow role, one daily routine; read its runs for a week before you hire it any colleagues.