Agents
An agent in QuoxCORE is a configured specialist: a system prompt, a set of tools, a mode, and an optional team. QuoxCORE ships 63 agents out of the box. You can configure custom agents and third-party agents (Claude Code, Devin, Hermes) using the same registration model.
Three modes
Every agent runs in one of three modes. The mode determines whether the agent loops, which tools it can access, and how much budget it can consume per invocation.
agent:standardThe agent uses a fixed system prompt and a fixed set of tools. It does not loop. One prompt in, one response out. Best for classification, summarisation, and structured extraction.
agent:agenticThe agent runs a think-act-observe loop. It selects tools, executes them, reads the result, and decides what to do next. The loop continues until the task is complete or the iteration limit is reached. Crystal retrieval and mid-loop drafting are active in this mode.
agent:browserThe agent controls a headless browser. It navigates, clicks, fills forms, and extracts data. The observation stream includes page DOM snapshots. Budget consumption is measured per page interaction. Requires the BrowserQ plugin.
The mode discriminator
Every agent definition contains a mode field. QuoxCORE reads this field before dispatching a task. The mode discriminator prevents an agentic loop from being invoked in a context that expects a single response, and vice versa. It is how the platform avoids runaway loops when a simple extract would suffice.
During the P3 Agents Unification work, the mode field was made the canonical runtime switch. Legacy agents without a mode field default to agent:standard.
Teams and team:invoke
Agents can be composed into teams. A team has a topology — star (one supervisor, many workers), hierarchical, pipeline, or mesh — and a routing policy. When a task exceeds a single agent's capability, the supervisor uses team:invoke to delegate sub-tasks to specialist agents.
Every delegation is logged in the AOCL evidence chain. You can see exactly which agent delegated to which, at what time, with what tool calls, and what the outcome was.
Human-in-the-Loop checkpoints can be inserted anywhere in a team workflow. When an agent reaches a HITL node, execution pauses and an inbox item appears for the approving user. The inbox correctly routes to the user's organisation — not the default fallback — since the HITL-ORGID-PLUMB fix shipped in P8.F.
Agents vs Assistants
In QuoxCORE, Assistants are supervisors: they hold conversation context, route tasks, and maintain persona continuity. Agents are workers: they execute discrete tasks, use tools, and return results. An Assistant may invoke many Agents across a single conversation. The two are not duplicates — they are at different levels of the hierarchy.