On this page
← All work

Agentic Task Management System

An LLM agent that plans and executes multi-step tasks using the ReAct framework and the Model Context Protocol (MCP) to interact with external tools.

Solo Project · Independent2025
  • LLM Systems
  • Agents
  • ReAct
  • MCP

A working agentic system built to explore how LLMs can move from single-shot answers to multi-step execution over real tools.

Problem

LLMs are good at generating text but weak at doing things. Turning an instruction like "collect the tickets, summarize them, and file a report" into tool calls, retries, and verification is a systems problem, not just a prompting problem.

Research Question

How should an agent be structured — reasoning loop, tool interface, and failure handling — so it can complete multi-step tasks reliably?

System

User instruction
        │
        ▼
┌───────────────────┐      tools         ┌──────────────┐
│ ReAct Agent       ├───────────────────►│ MCP Server    │
│ think → act → obs │                    │ (tools)       │
└───────────────────┘                    └──────────────┘
        │
        ▼
Task state / memory
  • ReAct loop: interleave reasoning (think), action (act), and observation.
  • MCP: a standardized tool interface so the agent can call external services without hard-coded wrappers.
  • State and retries: the agent keeps task state across steps and retries failed tool calls with bounded attempts.

My Contribution

  • Implemented the ReAct loop and prompt scaffolding.
  • Built an MCP server exposing task-management tools (search, create, update, summarize).
  • Added observation parsing, retry/backoff, and a simple task-state store.

Experiments

Compared task completion with and without the MCP tool layer, and with different numbers of reasoning steps allowed.

Results

Tool-augmented execution completed structured tasks that a plain chat model could not. Failure handling (retries and explicit error observations) materially reduced cascading failures.

Results are placeholder text — replace with actual measurements.

What I Learned

  • The tool interface design (MCP) matters more than the prompting details.
  • Explicit observations — feeding tool output back verbatim — keep the agent grounded.
  • Bounded retries and a hard stop prevent runaway cost on bad instructions.