AI Agents & Automation Internship

Build autonomous agents that use tools, not chatbots that answer questions.

The agent stack is where most of the interesting AI engineering work is happening, and almost no internship programme teaches it. You will build agents that plan, call real tools, recover from failure and run against real APIs — using the same patterns we use in production client work.

Intermediate4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Design a multi-step agent with explicit state and control flow
Give an agent real tools and handle tool failure safely
Understand when an agent is the wrong answer
Ship an agent behind an API endpoint

Who it suits

Students and early-career developers comfortable with Python who want the AI-engineering skill employers are actually short of.

What you'll work with

Python
LangGraph
LangChain
MCP
OpenAI/Claude/Gemini APIs
FastAPI

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Tool-Using Research Agent

~10h estimated

Build an agent that answers a research question by calling a search tool and citing its sources.

Define a small tool set (web search, fetch page, summarise). Use a loop or a LangGraph state machine so the agent decides which tool to call next, and stop it looping forever with a step budget.

Assessed against

At least two distinct tools the agent can choose between
A hard step limit so the agent cannot loop indefinitely
Every claim in the final answer is traceable to a fetched source
Graceful handling when a tool call fails or returns nothing

You'll finish with

GitHub repo with a README showing 3 example runs
A short recorded walkthrough

Multi-Agent Workflow with LangGraph

~12h estimated

Split a task across specialised agents that hand work to each other.

Model the workflow as a graph with explicit state: e.g. a planner, a worker and a reviewer. Make the reviewer able to send work back — that edge is the interesting part.

Assessed against

Three or more nodes with a clearly drawn state schema
At least one conditional edge (a real decision, not a straight line)
A diagram of the graph in the README
Explanation of what happens when the reviewer rejects twice

You'll finish with

GitHub repo
Architecture diagram
Written note on the failure mode you found

Build Your Own MCP Server

~12h estimated

Expose a real capability to an AI client through the Model Context Protocol.

Pick something with genuine side effects — a file store, a database query, a calendar. Implement the MCP server, then connect it to an MCP-capable client and drive it end to end.

Assessed against

At least two MCP tools with typed inputs
Input validation, so a bad argument cannot corrupt state
A clear boundary on what the server refuses to do
Setup instructions someone else can follow

You'll finish with

GitHub repo
Screen recording of a client using your server

Agent Reliability Harness

~10h estimated

Measure whether your agent actually works, instead of assuming it does.

Build an evaluation set of 20+ cases with known-good outcomes. Run the agent repeatedly and report success rate, average steps, token cost and the failure categories you observe.

Assessed against

20+ evaluation cases covering edge cases, not just happy paths
Reported success rate with the exact definition of success
Cost per successful run
A written analysis of the top three failure modes

You'll finish with

Evaluation repo + results table
Short write-up of what you would fix first

Free to apply and free to complete. The completion certificate is optional.

Other domains

Generative AI & LLMs

Go past prompt tinkering into retrieval, evaluation and cost control.

Machine Learning

Classical ML done properly — including the evaluation everyone skips.

Data Science

Turn a messy dataset into a decision somebody can act on.

Ready to Transform Your Enterprise with AI?

Book a free 30-minute consultation with our AI architects. Discover how agentic AI can automate your workflows and multiply your team's output.

No commitment required · Free architecture review · NDA protected

AI Agents & Automation Internship

Build autonomous agents that use tools, not chatbots that answer questions.

Intermediate4 weeks · remote4 briefs · complete any 2Free to join

What you'll be able to do

Design a multi-step agent with explicit state and control flow

Give an agent real tools and handle tool failure safely

Understand when an agent is the wrong answer

Ship an agent behind an API endpoint

Who it suits

Students and early-career developers comfortable with Python who want the AI-engineering skill employers are actually short of.

The project briefs

Complete any 2 of these 4to finish the programme. Assessment criteria are published here so you know what “good” means before you start.

Tool-Using Research Agent

~10h estimated

Build an agent that answers a research question by calling a search tool and citing its sources.

Define a small tool set (web search, fetch page, summarise). Use a loop or a LangGraph state machine so the agent decides which tool to call next, and stop it looping forever with a step budget.

Assessed against

At least two distinct tools the agent can choose between
A hard step limit so the agent cannot loop indefinitely
Every claim in the final answer is traceable to a fetched source
Graceful handling when a tool call fails or returns nothing

You'll finish with

GitHub repo with a README showing 3 example runs
A short recorded walkthrough

Multi-Agent Workflow with LangGraph

~12h estimated

Split a task across specialised agents that hand work to each other.

Model the workflow as a graph with explicit state: e.g. a planner, a worker and a reviewer. Make the reviewer able to send work back — that edge is the interesting part.

Assessed against

Three or more nodes with a clearly drawn state schema
At least one conditional edge (a real decision, not a straight line)
A diagram of the graph in the README
Explanation of what happens when the reviewer rejects twice

You'll finish with

GitHub repo
Architecture diagram
Written note on the failure mode you found

Build Your Own MCP Server

~12h estimated

Expose a real capability to an AI client through the Model Context Protocol.

Pick something with genuine side effects — a file store, a database query, a calendar. Implement the MCP server, then connect it to an MCP-capable client and drive it end to end.

Assessed against

At least two MCP tools with typed inputs
Input validation, so a bad argument cannot corrupt state
A clear boundary on what the server refuses to do
Setup instructions someone else can follow

You'll finish with

GitHub repo
Screen recording of a client using your server

Agent Reliability Harness

~10h estimated

Measure whether your agent actually works, instead of assuming it does.

Build an evaluation set of 20+ cases with known-good outcomes. Run the agent repeatedly and report success rate, average steps, token cost and the failure categories you observe.

Assessed against

20+ evaluation cases covering edge cases, not just happy paths
Reported success rate with the exact definition of success
Cost per successful run
A written analysis of the top three failure modes

You'll finish with

Evaluation repo + results table
Short write-up of what you would fix first

Free to apply and free to complete. The completion certificate is optional.