AI Agent Projects

Overview

Agents are where LLMs stop being autocomplete and start being useful. The question isn't "can the model generate text?" — it's "can the model take actions in the world, recover from failures, and achieve a goal across multiple steps?"

This collection documents my experiments building agents that go beyond toy demos.

Agent Architectures Explored

ReAct (Reasoning + Acting)

The classic pattern. The model alternates between thinking steps ("I need to look up X") and action steps (calling a tool). Implemented from scratch without LangChain first to understand what's really happening, then rebuilt with LangChain for production use.

Key finding: The quality of tool descriptions matters more than the tool implementations. Vague tool descriptions cause the model to hallucinate tool signatures.

---

Tool-Use Agent with Structured Outputs

Built an agent that browses local file systems, reads CSV data, generates Python code to analyse it, executes the code via a sandboxed subprocess, and returns a summary — all in one chain.

The critical component is structured output: the model must return valid JSON for every action, not free text. Using Pydantic models as the output schema drops hallucinated tool calls to near zero.

---

Memory Patterns

Explored three memory patterns:

Buffer memory: Simple sliding window of last N messages. Works for short tasks, falls apart on anything requiring long-horizon context.

Summarisation memory: Compress old context into a running summary. Loses precision but scales. Good for conversational agents.

Entity memory: Extract and store named entities (people, projects, dates) as a structured knowledge base. The agent can retrieve specific facts without loading full history. Best for domain-specific assistants.

A Real Use Case: Research Assistant Agent

Built a research assistant that:

1. Accepts a research question
2. Decomposes it into sub-questions
3. Searches a local knowledge base (PDF documents via FAISS)
4. Synthesises findings across sources
5. Returns a structured report with citations

This runs end-to-end in FastAPI, accepts questions via a REST endpoint, and streams the response via SSE.

What Breaks Agents

• Token limits: Long tool outputs blow the context window. Need aggressive summarisation between steps.
• Reward hacking: The model finds the easiest path to a "done" state, not necessarily the correct one. Need explicit verification steps.
• Error recovery: Most frameworks don't teach the model to retry intelligently after tool failures. Had to add custom retry logic with error context fed back into the prompt.

What's Next

• Multi-agent coordination: one planner, multiple specialist sub-agents
• Persistent memory across sessions using a vector store
• Evaluating agent trajectories, not just final outputs