Agents
The specialized agents behind a flow and the safeguards that keep them on track.
A flowflowOne run of PentAGI: a goal, the tasks created from your messages, and everything the agents did along the way. is carried out by a team of agentsagentAn AI agent of PentAGI: a model with its own prompt, settings and tools that does one kind of work in a flow, such as testing, writing code or searching the web.. Each agent typeagent typeThe role an agent plays, such as pentester, coder or searcher. Model settings are chosen per agent type. has its own prompt, model settings and toolstoolWhat an agent acts through: a built-in capability such as the terminal, the isolated browser or web search, or a program in the sandbox such as nmap, metasploit or sqlmap., so you can assign stronger or cheaper models per role.
Agent types
| Agent | Role |
|---|---|
primary_ | Main reasoning agent that drives a flow |
assistant | Interactive assistantassistantThe interactive mode: you work with an agent in a chat and decide on the next step together. mode |
generator | Splits each tasktaskOne request in a flow: its first message, or a later message that does not answer an agent's question. PentAGI splits each task into subtasks. into subtaskssubtaskOne step of a task. PentAGI plans the subtasks of each task and adjusts them as the work goes on. |
refiner | Refines the plan as results come in |
adviser | Expert advice; mentor in execution monitoring and planner in task planning |
reflector | Corrects agents that stop producing tool callstool callA request from the model to run one of the agent's tools, such as a terminal command or a web search. |
searcher | Information gathering and web search |
enricher | Gathers extra context from memorymemoryWhat agents keep so they can build on earlier work instead of starting from scratch: long-term, working and episodic memory. and the sandboxsandboxThe isolated Docker container in which agents run their tools. The interface shows it as Terminal. before the adviser answers |
coder | Writes and analyzes code |
installer | Installs and sets up tools in the sandbox |
pentester | Penetration testing and security assessment |
simple, simple_ | Plain and JSON-structured completions |
Two more agents work behind the scenes: the memorist answers other agents' questions about past work from long-term memory, and the reporter writes up the result of each task.
How work is delegated
Safeguards
- Tool call limits (always on) —
MAX_(default 100) caps the model turns in the chains ofGENERAL_ AGENT_ TOOL_ CALLS primary_,agent assistant,pentester,coderandinstaller, andMAX_(default 20) caps them for the other agents. One turn can call several tools. Values from 1 to 5 are raised to 6, and 0 or less means the default.LIMITED_ AGENT_ TOOL_ CALLS - Repeat detection (always on) — from the third call in a row to the same tool with the same arguments, the tool is not run again: the agent gets the previous answer back with a request to change course. The seventh such call ends the chain.
- Reflector (always on) — steps in the first time a model answers without a tool call and guides the agent back to one; after three unsuccessful rounds the chain stops.
- Execution monitoring (beta,
EXECUTION_) — forMONITOR_ ENABLED primary_,agent pentester,coder,installer, andassistantwhen it works with agents, the adviser acts as a mentor when the agent calls the same toolEXECUTION_times in a row (default 5) or makesMONITOR_ SAME_ TOOL_ LIMIT EXECUTION_tool calls since the last advice (default 10). The trigger counts calls rather than detecting a stall; the adviser then reviews the agent's progress and recommends whether to continue, change approach or stop.MONITOR_ TOTAL_ TOOL_ LIMIT - Intelligent task planning (beta,
AGENT_) — the adviser drafts 3–7 concrete steps before specialists start a subtask.PLANNING_ STEP_ ENABLED
Execution monitoring and task planning roughly double result quality for models under 32B parameters, at the cost of 2–3× more tokens and time. They are off by default.
In Enterprise, the planner always runs and execution monitoring is on by default with limits of 20 and 40 — see Agent profiles.
Last updated on