Build on AgentScape
Build agents your company can actually trust
No prompt spelunking, no YOLO deploys. You describe the job, prove the behavior in a test loop, approve it for your team, and run it with budgets and permissions enforced by the platform.
From idea to approved agent
01
Describe the job
In the dashboard's Create Agent tab you write the spec, not the prompt: intended outcome, workflow steps, decision rules, prohibited uses, tone, and output format. AgentScape compiles it into a governed system prompt.
02
Prove it in the test loop
Every agent gets a guided test bench built from its own spec — happy path, missing input, prohibited request, escalation edge, and messy input. Failed runs explain themselves: access, budget, endpoint, validation, or runtime.
03
Approve it for your company
New agents are drafts. Someone with approval permission signs off — and approval is blocked until the five test cases have actually run. Approved agents join your company's list, governed by roles you control.
04
Run it anywhere
Run governed jobs from the dashboard with budgets, tool grants, and full run history — or pair the Windows desktop app and run the same signed agents locally with permission prompts.
Guardrails are the product
Everything a builder ships inherits the company's governance — the same controls you see on the Company page.
Draft-first lifecycle
Nothing your team writes can run for the company until it's tested and approved. Rejected or suspended agents stay visible to their owners, never runnable.
Budgets that bite
Token, cost, runtime, and tool-call ceilings per run, per day, and per month — enforced before a run starts, not tallied after.
Scoped tool grants
Agents call tools only through explicit grants with scopes, rate limits, and per-run call caps. No grant, no call.
Roles and permissions
Owners decide who can build, test, approve, install, and run — per company, from the Company page. Policy changes apply to every member and paired device immediately.
Ready to go public?
Push a company-approved agent to platform review to list it in the public directory — verification covers publisher identity, signed manifests, and endpoint checks. Running an agent elsewhere already? Register the listing directly.