AI Agent Autonomy (Infographic)

Description: At the basic level, an LLM agent acts as a "tool user." Developers predefine a set of functions (tools) the agent is allowed to call. The LLM's job is to understand the request, pick the most suitable tool from the library, and generate the correct parameters to call it.

Example: A weather agent has a `get_weather(city)` tool. Given the request "What's the weather in London?", the LLM selects this tool and calls it with the parameter `city='London'`. The agent doesn't write any code — it just uses what already exists.

Description: This approach, used by Anthropic, distinguishes between "agents" and "workflows." An agent is an LLM orchestrator that dynamically breaks a complex task into subtasks and delegates them to specialized worker agents. It's similar to a team lead distributing work across a team.

Architecture: A central model (the orchestrator) coordinates the work, while worker models carry out specific operations using predefined tools. This improves reliability and simplifies debugging, but it still doesn't involve generating new code "on the fly."

Description: This is where the qualitative leap happens. The agent doesn't just use existing functions — it can *create new ones* at runtime. Given a task with no ready-made tool for it, the LLM generates a block of code (in Python, for example) to solve it.

Safety: This is a key aspect. The generated code runs in a strictly isolated, secured environment (a Docker container, for example) with minimal permissions, to prevent any risk to the core system. It's a "sandbox" where the AI can safely experiment.

Description: This is the cutting edge of research. "Self-improving coding agents" can not only generate code for external tasks but also autonomously edit and improve *their own* underlying code. This is a meta-level of autonomy.

Potential: Such an agent can optimize its tool usage, find more efficient prompting schemes, or even fix bugs in its own logic. This opens the door to systems that evolve and improve over time without direct human intervention — a step toward AGI (Artificial General Intelligence).

• ✓Hallucinations: Generating plausible-sounding but factually incorrect or fabricated information. This directly affects code correctness.
• ✓Solving complex problems: Difficulty with tasks that require deep domain knowledge, innovation, or handling ambiguous requirements.

• ✓High complexity: Generated code can be needlessly complex, making it harder to understand, debug, and maintain long-term.
• ✓Suboptimal solutions: The code can be inefficient, reflect outdated practices, or be "overfitted" to specific data.

• ✓Unintentional vulnerabilities: Risk of introducing vulnerabilities such as hardcoded credentials, SQL injection, or buffer overflows.
• ✓Reproducing bad practices: Models can reproduce unsafe coding practices present in their training data.

• ✓Amplifying bias: Reproducing and amplifying biases already present in the data, leading to unfair outcomes.
• ✓Accountability and control: Difficulty in assigning accountability for errors, and the need for human oversight in critical scenarios.
• ✓Goal misalignment: The risk that an agent's goals diverge from the user's well-being, leading to manipulation or misuse.