Autonomous AI Agents
What AI Agents are in Aimogen Pro, how they differ from a chatbot with tools, and the execution model that runs them step by step.
An AI Agent is given a task, not a prompt. It works out the steps itself, calls tools to carry them out, observes the results, and continues until the task is done or its step budget runs out.
Added in 2.7.7.
Agents, chatbots and OmniBlocks#
Three different things, often confused:
| Chatbot | OmniBlock | Agent | |
|---|---|---|---|
| Who decides the steps | The visitor, one message at a time | You, in advance | The AI |
| Runs | Interactively | On a trigger or schedule | In the background, over many steps |
| Uses tools | Optionally, one per reply | No — blocks are the steps | Yes, repeatedly |
| Good for | Answering questions | Repeatable known processes | Open-ended tasks |
| Predictable | Per reply | Completely | Least |
Use an OmniBlock when you can write down the steps. It is cheaper, faster and predictable.
Use an agent when you cannot — "audit the site for posts with no featured image and fix them", "research this topic and draft a briefing", "find and repair broken internal links".
The execution loop#
- The agent receives a task.
- If planning is enabled, it produces a plan first.
- It decides on an action — call a tool, or conclude.
- Aimogen Pro executes the tool and returns the result.
- The result goes into the agent working memory.
- Back to step 3, until the task is complete or
Max Stepsis reached.
Every step is recorded in the aiomatic_agent_runs table, giving a full audit trail of prompts,
responses, tool names, parameters and results.
Execution is asynchronous, driven by the aiomatic_agent_run_single cron event, with
aiomatic_agent_recover and aiomatic_agents_watchdog handling stalls. See
Running and monitoring agents.
In this section#
Where agents live#
Aimogen Pro › AI Agents. Agents are stored as aiomatic_agent posts, with configuration and state in
post meta:
| Meta key | Holds |
|---|---|
_agent_config | The configuration |
_agent_status | Current status |
_current_task | The task being worked on |
_agent_current_step | Step counter |
_agent_memory | Working memory |
_agent_recover_count | Recovery attempts |
_last_activity | Last activity timestamp |
_model | The model in use |
A set of example agents is imported on first activation from the plugin defaults/ folder.
Realistic expectations#
Agents are impressive when they work and frustrating when they do not. What experience suggests:
They are good at well-bounded tasks with clear success criteria, work involving several tools, and tasks where the exact sequence depends on what is found.
They are less good at long open-ended tasks (they drift), anything needing judgement about your business, and tasks where a mistake is expensive.
They cost more than you expect. Budget for several times the token spend of a single generation.
Start with a narrow task, a small tool set and a low step limit. Widen only once you have seen it succeed.
Related#
- Chatbot Extensions — the same tools, in a conversation
- OmniBlocks — for processes you can specify
Still stuck? Open a support ticket and include the diagnostics from Aimogen Pro › System & Logs › System Info.