01 Agentic Systems 3 min read 646 words
Agent Fundamentals (Dec 2025)
Agents are LLM-powered systems that move beyond "chat" into "autonomous problem solving." In late 2025, the definition of an agent has shifted from simple ReAct loops to Closed-Loop Reasoning Systems that utilize built-in "System 2" thinking.
The Agent Formula
Modern agency is often described as:
Agent = Reasoning Model + Tool Use + Persistent Memory + Environment Feedback
2025 Nuance: In 2023, agents were "wrappers." In 2025, agents are increasingly Integrated. Frontier models (like OpenAI o1 or DeepSeek R1) have the "Thinking" process baked into the pre-training, making the agent loop more stable and less prone to "stalling."
System 1 vs. System 2 Thinking
Architecting an agent requires choosing the right "Thinking Mode":
| Mode | Cognitive Type | Analogy | 2025 Stack |
|---|---|---|---|
| System 1 | Fast, intuitive, reactive | Reflexes | GPT-4o / Sonnet 3.5 |
| System 2 | Slow, logical, planning | Deliberation | o1 / R1 / Reasoning Loops |
The Design Pattern: Use System 1 models for "Fast UI" and "Routing." Use System 2 models for "Decision Gates" and "Complex Planning."
Agency Levels
Not every autonomous system is an "Agent." We categorize them by the Level of Agency:
- L0: Scripted Chains: Fixed sequence (e.g., standard LangChain).
- L1: Tool-Enabled: Model picks a tool but doesn't plan.
- L2: ReAct Agent: Simple loop of "Thought -> Action -> Observation."
- L3: Autonomous Planner: Decomposes a goal into a graph of sub-tasks.
- L4: Ambient Agent: Runs in the background, intervenes only when necessary.
Core Components
1. The Reasoning Model (The Executive)
The CPU of the agent. It determines the "Path to Success."
2. Tools (The Limbs)
Interfaces (APIs, Browsers, DBs) that allow the agent to affect the world.
3. Memory (The Experience)
- Short-term: Context window (KV Cache).
- Long-term: Vector DBs or persistent state (e.g., Mem0).
The Agent Lifecycle
- Intake: Receive user goal.
- Decomposition: Break goal into sub-steps.
- Execution: Call tools and handle results.
- Reflection: Evaluate if the observation got the agent closer to the goal.
- Completion: Synthesize final proof for the user.
Interview Questions
Q: Why is a "Reasoning Model" (like o1) better for agency than a standard LLM?
Strong answer: Standard LLMs (System 1) predict the very next token based on pattern matching. When they encounter an error in a tool call, they often hallucinate a fix instead of admitting the failure. Reasoning Models use Chain-of-Thought (CoT) during inference. They "think" through multiple hidden turns before outputting a response. For an agent, this means higher Path Reliability—the model is significantly less likely to enter an infinite loop or try the same failing action twice because it has already simulated the failure internally.
Q: How do you prevent "Agentic Drift" in long-running tasks?
Strong answer: Agentic Drift occurs when the sub-steps take the agent so far from the original goal that it loses context. In 2025, we solve this with Goal Anchoring. We include the "Original Objective" as a pinned system message and use a Secondary Observer Model (a smaller, cheaper model) to score every agent action against the original objective. If the score drops below a threshold, the agent is forced to "re-plan" from the root.
References
- Kahneman, D. "Thinking, Fast and Slow" (applied to AI, 2025)
- OpenAI. "Learning to Reason with LLMs" (2024)
- DeepSeek. "R1: Cold-Start Data for Reasoning" (2025)
Key takeaways
01
Five levels of agency, not a binary
L0 scripted chains, L1 tool-enabled, L2 the ReAct loop, L3 an autonomous planner decomposing into a task graph, and L4 ambient agents that run in background and intervene only when needed.
02
Match the thinking mode to the step
The page routes fast System 1 models to UI and routing work and slow System 2 reasoning models to decision gates and complex planning; the split is per-step, not per-application.
03
Reasoning models improve path reliability
Because they simulate a failure internally before answering, they are less likely to retry the same failing tool call or spin in a loop than a model matching the next token.
04
A cheaper observer catches drift
Goal anchoring pins the original objective as a system message while a small observer model scores each action against it, forcing a re-plan from the root when the score drops.
05
Reflection is the stage people skip
The five-stage lifecycle is intake, decomposition, execution, reflection, completion; reflection asks whether the last observation actually moved the agent closer to the stated goal.