Designing Progressive Large‑Model Agents: Architecture, Frameworks, and Real‑World Practices

Background

Agents were introduced to mitigate large‑model hallucinations, reduce wasted inference compute, and shift from single‑turn question answering to goal‑driven autonomous execution.

Development Stages of Agents

Stage 1 – Prompt engineering + tools : Models learn to invoke external tools (e.g., calculators, search engines) via prompts.

Stage 2 – Planning + memory : Short‑term memory supports context tracking; long‑term memory stores accumulated experience.

Stage 3 – Multi‑agent collaboration : Frameworks such as CrewAI and AutoGPT coordinate multiple specialized agents as a small team.

Stage 4 – Progressive agents : Designs like DeepMind AlphaEvolve perform self‑optimization, evolution, and composition across multiple rounds.

Agent Framework Categories

Workflow‑type (fixed pipelines for structured tasks) : Example – LangGraph (built on LangChain) uses a node + state + edge graph model to support loops, conditional branches, and long dialogues.

Collaboration‑type (multiple agents cooperate) : Example – CrewAI focuses on building collaborative multi‑agent systems.

Research/Evolution‑type (search, evolution, self‑optimization) : Examples – DeepResearch (Perplexity) and AlphaEvolve implement a generate‑evaluate‑select‑optimize loop for autonomous coding agents.

Core Components of an Agent

Perception : Understand environment and inputs (user queries, external data).

Memory : Short‑term memory supports immediate context; long‑term memory retains experience.

Planning : Decompose complex tasks into subtasks.

Tools : Invoke APIs, databases, code executors, etc.

Action/Executor : Carry out the plan and deliver results.

Reflection : Review the process to improve stability and robustness.

Manus, a general‑purpose ReAct‑style agent, illustrates these components with:

KV‑Cache design that appends context rather than replacing it for identical prefixes.

Constrained decoding that explicitly disables certain tools while preserving error cases.

File‑system‑based memory that avoids irreversible compression of context.

TODO‑driven bias correction that reduces goal drift by restating and updating a TODO.md file.

Progressive Agent Loop

Task decomposition : Break a complex problem into subtasks, optionally using a Task Decomposer module powered by LLMs or rules.

Solution generation : A Generator Agent produces candidate answers, code, or plans.

Evaluation & reflection : A Critic Agent or reward function scores the candidate; failures are recorded.

Iterative improvement : Prompt, tool usage, or plan is adjusted based on feedback; a new solution is generated.

Convergence & output : When a solution meets predefined quality criteria or iteration limits, the final result is emitted.

This closed‑loop evolutionary system mirrors AlphaEvolve’s “Generate → Evaluate → Evolve” cycle.

Implementation Building Blocks

Prompt sampler : Assembles prompts for each question.

LLMs ensemble : Selects and combines appropriate models.

Evaluator : Assesses execution results using rules, LLMs, or reinforcement learning.

Program database : Stores generated programs; an evolutionary algorithm decides which programs are reused in future prompts.

Typical Implementations

Research / knowledge exploration : A Retrieval Agent queries external databases or literature each iteration; a Critic Agent checks coverage and credibility; the loop gradually builds a research report.

Code agent (e.g., Cursor) : Generator creates code; Executor compiles and runs it; Evaluator collects error messages; Refiner iterates on the code until tests pass.

Complex task scheduling (supply chain / operations) : Planner splits tasks into standard nodes; Executor carries out each node while writing to Memory; Evaluator validates node results (e.g., inventory, logs) and retries on failure; Orchestrator oversees overall progress.

References

https://www.promptingguide.ai/zh

https://lilianweng.github.io/posts/2023-06-23-agent/

https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/