Skip to main content
A Condor Agent is a folder on disk that Condor runs as a tick loop. On each tick it reads the market, asks an LLM what to do, and turns the decision into trades through Hummingbot executors or controller-based bots. A risk gate checks every call that could move money.

Design philosophy

Condor Agents separate what LLMs do well, reasoning under uncertainty, from what traditional software does well, reliable and repeatable execution. The result is an agent that thinks like a discretionary trader but acts like a systematic one, with a full record of every tick.

Executor-based trading

Agents never place orders directly. The risk gate always refuses place_order. An agent trades in one of two ways:
  • Executors. It creates Hummingbot executors (position, grid, DCA, order or LP) with the create_*_executor tools.
  • Controller-based bots. It deploys and retunes a Hummingbot bot that runs a controller, with manage_bots. A deploy may declare a loss cap (max_global_drawdown_quote). Leaving it out deploys with no platform-side cap. A cap you set can’t exceed the session’s max_position_size_quote, and must be below the capital the bot trades. See the Market Making Expert for a loop that works this way.

Ownership through controller_id

Each loop session has its own ID, such as my_agent.my_loop_3. Every executor the session creates must carry that ID as its controller_id; the risk gate refuses a create without it. This gives you:
  1. Isolation. An agent only manages its own executors. list_executors returns the whole fleet on the server, so the agent can see other executors, but the risk gate refuses any attempt to stop an executor owned by another live session.
  2. A virtual portfolio. Positions and PnL are tracked per controller_id, so each session has its own book even when several agents share one exchange account. See Inventory.
  3. Position handover. When an executor closes with keep_position=true, the position stays in that session’s inventory.
When you start a loop again, the new session adopts the executors an earlier, stopped session of the same loop left running, and can manage and stop them. See Sessions.

Mental model

An agent is a folder on disk and a tick loop in memory. The folder is its long-term memory; the data it gathers each tick is its situational awareness; the LLM is its decision function; executors and bots are its hands; the risk gate is the wrist it can’t move past.

Run modes

Choose the mode in the Start dialog or with the loop’s execution_mode setting. Only loop creates a session. Start with dry_run to check that the agent, server and exchange are wired up.

Risk limits

Each session has six risk limits, set in the loop’s default_config.risk_limits or in the Start dialog: What each one does is on the Architecture page.

Stop vs. shutdown

  • Stop ends the loop and leaves its executors, positions and bots open.
  • Emergency shutdown closes the session’s executors according to the agent’s shutdown.md policy, then stops. A shutdown_drawdown_pct breach triggers it automatically.
See Stopping a loop.

Getting started

Learn more