Agents that ship work, not just chat.
SoxAI ships ready-made platform agents every tenant can use — and lets you build custom agents with model, system prompt, tool set, and per-turn cost caps. Lead + Worker orchestration turns one request into a coordinated multi-agent task.
From a single prompt to a finished task
Six primitives that turn an LLM call into an agent that actually completes work — with the guardrails enterprises need.
Platform agents, published once
Platform agents belong to the deployment, not any tenant. Publish one and every tenant's users can use it immediately — no per-tenant rebuild. Mark one as the default users land on.
Custom agents, fully configurable
Build your own agent: pick the model, write the system prompt, choose the tool set, set a per-turn cost cap, and expose it as a called-as endpoint. Editing a published agent updates what every user gets.
Lead + Worker orchestration
A Lead agent plans and spawns Worker sub-agents to run multi-step work in parallel. Workers call tools, return results, and the Lead assembles the final answer — true multi-agent orchestration, not a single chat loop.
Tool-calling with full audit
Agents call real tools — search the codebase, run tests, open PRs. Every tool call is logged with its arguments, result, and cost. Per-turn cost caps prevent runaway spend before it happens.
Visibility controls
Keep an agent in draft (platform only) while you tune it, then flip to published when ready. Optionally mark one agent as the default so users land on it when they open the chat.
Cost caps per turn
Set a hard cost ceiling per agent turn. When a turn approaches the cap, the agent stops and reports — no surprise bills from a tool-calling loop that refused to quit.
The agent console
Publish platform agents, configure custom ones, and watch Lead + Worker orchestration run in real time — all from the admin console.
Lead + Worker, explained
A Lead agent receives the request, breaks it into steps, and spawns Worker sub-agents. Each Worker runs a focused sub-task — research, code, SQL — calling tools as needed. Results flow back to the Lead, which assembles and ships the final output.
Every step is auditable: which Worker ran, which tools it called, what it cost, and whether it succeeded. Orphaned workers are cleaned up automatically — a running job whose session vanished is reclaimed, not left heartbeating forever.
Breaks the task into 3 sub-tasks
search_codebase(), read_file()
run_tests(), open_pr()
Put your agents to work
Publish a platform agent in minutes, or build a custom one with full tool-calling and cost guardrails.
Get started