Agentic behaviors
Also written: Autonomous AI, AI agents, Agentic exploration, Tamper With Their Own Traces
AI that can act on its own to plan and complete multi-step tasks, often using tools.
Agentic behaviors describe AI systems that can act independently to achieve goals. Instead of just responding to a prompt, these AI agents can plan a series of steps, execute them, and even use tools or other AI models to get the job done.
For example, an agentic AI could be asked to plan a vacation. It might then autonomously search for flights, compare hotel prices, book accommodations, and create an itinerary, all without further instructions for each step.
Agentic behaviors refer to the capability of AI systems to autonomously plan, execute, and coordinate multi-step actions to achieve predefined goals. This often involves the AI interacting with its environment, utilizing external tools (like web browsers or APIs), or even orchestrating other AI models. The key distinction is the AI's ability to self-direct its actions over a sequence of operations, rather than simply responding to a single prompt.
Tradeoffs include increased complexity in development and potential for unintended consequences if not carefully controlled. While the concept has been explored for years, recent advances in large language models have made more sophisticated agentic systems feasible.
Recent news highlights the growing capabilities of AI systems to act autonomously and pursue complex goals. This is leading to discussions about their potential impact on various sectors, as well as the legal and ethical considerations surrounding their deployment. The development of platforms to test and evaluate these autonomous systems further underscores their increasing relevance.
- OpenAI hit with landmark lawsuit following Hugging Face hackOpenAI · Axios AI ·
- LLM Agents Can Easily Tamper With Their Own TracesAnthropic · arXiv cs.AI ·
- Teaching Everyone to Fish for TokensAnthropic · Interconnects (Nathan Lambert) ·
- When millions of AI agents meetGoogle · Google DeepMind: The Podcast ·
- What it feels like to work with MythosAnthropic · One Useful Thing (Ethan Mollick) ·
- SHADE-Arena: Evaluating Sabotage and Monitoring in LLM AgentsAnthropic · Anthropic Research ·
- Anthropic Claude
- Anthropic's work includes evaluating AI agents for sabotage and monitoring in systems like Claude 5 Fable, which demonstrates advanced task handling.
- Google Gemini
- Google Gemini is exploring the implications of millions of AI agents interacting and their potential to reshape science, economies, and security.
- OpenAI ChatGPT
- OpenAI faces legal scrutiny over AI agents acting beyond control, indicating a growing concern about the autonomy and oversight of these systems.
- Prompt engineering
- This is the art of crafting inputs to guide an AI's response
- Reinforcement learning
- While RL can be used to train agents, agentic behavior d
- Task decomposition
- This is the process of breaking down a complex task into sma
The concept of AI agents acting autonomously has roots in early AI research, particularly in areas like intelligent agents and robotics. The term gained significant traction with the rise of sophisticated AI models capable of complex reasoning and tool interaction.
Ask your AI assistant to plan a weekend trip to a nearby city. Instruct it to find three potential activities, check their opening hours, and suggest a possible itinerary. Observe if it can autonomously perform these steps and present a coherent plan.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- WorkflowBuild an AI agent evaluation pipeline in one hourCompress weeks of AI agent testing into one hour using open-source tools and automated pipelines. · 60 min
- HeuristicAI agents need time-awareness for runtime controlAI agents require a sense of time to predict, estimate, and control their own actions and execution duration. · 5 min
- Anti-patternUsing prototype defaults for production AI agent architectureDon't let prototype settings for AI models and prompts become production defaults; they are not optimized for scale or cost. · 60 min
- Cost tipLeverage prompt caching to reduce AI agent token costsReuse previously processed prompt prefixes to avoid reprocessing and save on token costs for agents. · 30 min
- HeuristicComposable controls for agent autonomy and safetyUse composable controls that perform single jobs, enabling agents to run with autonomy at appropriate safety levels. · 10 min
Questions people ask
- Can any AI perform agentic behaviors?
- Not all AI systems exhibit agentic behaviors. It requires specific architectures and training that enable autonomous planning and action execution, often involving tool use.
- What's the difference between an AI assistant and an AI agent?
- An AI assistant typically responds to direct commands, while an AI agent can take initiative, plan, and execute tasks autonomously to achieve a goal.
First seen on Subvolts Jun 25, 2025. Updated Oct 12, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.