02 Agentic Systems 3 min read 634 words

Reasoning Loops: ReAct and Beyond (Dec 2025)

Reasoning Loops define the control flow of an agent. While ReAct was the 2023 baseline, late 2025 systems use more sophisticated patterns like Plan-and-Solve, Self-Reflexion, and Inference-Time Scaling.

agentsreasoningorchestrationcore
01loop

The Evolution of the Loop

EraPatternCore Philosophy
2023ReActInterleave thought and action.
2024ReflexionEvaluate errors and re-try.
2025System 2 LoopsUse hidden CoT for robust multi-step logic.
02+ acting

ReAct: Reasoning + Acting

The fundamental loop for 90% of agents:

  1. Thought: "I need to find X."
  2. Action: search_engine("X")
  3. Observation: "X is at Y."
  4. Repeat.

2025 Critique: ReAct is fragile. If the search returns "No results," a naive ReAct agent will often try the same search again. Modern loops inject "Negative Constraints" (e.g., "Don't try search results we've already seen").

03loops

Self-Reflexion Loops

Reflexion adds a "Critic" step to the loop.

The 2025 Benefit: By storing these "Reflections" in short-term memory, the agent builds a "Mental Map" of what doesn't work during the current session.

04plan-and-solve

Plan-and-Solve

Instead of deciding one step at a time (greedy approach), the agent creates a Static Plan first, then executes it.

  1. Planner: "I will do A, then B, then C."
  2. Executor: Carries out the steps.
  3. Re-planner: If step B fails, trigger a full re-plan rather than a local fix.

Why?: Planning reduces "Stochastic Errors." By committing to a path, the model is less likely to get distracted by noisy tool results.

05engineering langgraph

Flow Engineering (LangGraph)

In late 2025, we have moved from "Chat interfaces" to "State Machines."

  • Cyclic Graphs: Instead of a linear sequence, we define a graph where the model can loop back to a "Cleaning" node or a "Validation" node multiple times.
  • Micro-Agents: Each node in the graph is a specialized "Prompt" or "Tool."

Key Nuance: The "Agent" is no longer just the LLM; the agent is the Graph Execution Engine.

06questions

Interview Questions

Q: When would you use a "Reasoning Loop" (ReAct) vs. a "Plan-and-Solve" architecture?

Strong answer: I choose ReAct for Exploratory tasks where the environment is unpredictable (e.g., browsing a new website where you don't know the URL structure yet). The agent needs to react to every observation. I choose Plan-and-Solve for Predictable but complex workflows (e.g., generating a financial report from 5 known APIs). Planning prevents the model from "meandering" and allows for better parallelization of steps that don't depend on each other.

Q: What is "Inference-Time Scaling" and how does it relate to Agentic Loops?

Strong answer: Inference-Time Scaling (often associated with OpenAI's o1) refers to spending more compute during the response generation rather than just during training. In an agentic context, this means the model doesn't just output the first valid-looking action. It uses a Search Tree (like Monte Carlo Tree Search) to simulate different action paths internally before committing to the one most likely to succeed. This reduces the number of "Real World" tool calls needed, saving external API costs and reducing failure rates.

07references

References

  • Yao et al. "ReAct: Synergizing Reasoning and Acting" (2022/2025 update)
  • Shinn et al. "Reflexion: Language Agents with Iterative Homeostatic Learning" (2024)
  • Wang et al. "Plan-and-Solve Prompting" (2023)

Next: Tool Use and the Model Context Protocol (MCP)

summary · added by this rebuild

Key takeaways

01

ReAct repeats its own dead ends

A naive thought-action-observation loop re-runs a search that already returned nothing; modern loops inject negative constraints naming what has already been tried.

02

Reflexion adds a critic to the loop

An evaluator routes success to output and failure back through a reflection step, and storing those reflections builds a within-session map of what does not work.

03

Plan first when the environment is predictable

Plan-and-Solve commits to A then B then C and re-plans wholesale on failure, reducing stochastic drift; ReAct suits exploratory work where each observation must steer.

04

The graph is the agent

Flow engineering makes each node a specialised prompt or tool and lets cycles revisit a cleaning or validation node; control flow lives in the execution engine.