Thorsten Ball

Thorsten Ball is a software engineer, writer, and co-founder and co-creator of Amp. His AI work includes teaching developers how a code-editing agent operates and explaining the Amp team's experiments with tools and delegation. In his April 2025 tutorial, he distilled the agent's core to a short statement: “It’s an LLM, a loop, and enough tokens.”[2][6][3]

In How to Build an Agent, published April 15, 2025, Ball walks readers through a Go implementation using the Anthropic SDK and Claude 3.7 Sonnet. The program maintains conversation history locally, sends it to the model, executes requested tools, and returns their results for another response. He adds three tools: read_file, list_files, and edit_file. This turns a terminal chat program into one that can inspect and change files, without requiring a separate hand-written workflow for every task.[3]

Ball's worked example asks the agent to create a JavaScript FizzBuzz program, then change its limit from 100 to 15. The agent reads the file, edits the function call, and updates the comments to match. He presents the tutorial implementation as achievable in fewer than 400 lines, mostly boilerplate. He distinguishes that small inner loop from the additional editor integration, prompting, feedback, interface, and tooling the Amp team built—work he describes as “practical engineering and elbow grease.”[3]

Ball's June 10, 2025 note, Agents for the Agent, describes how model behavior changed the team's approach to subagents. Amp already had a read-only search agent. Other specialized agents had proved unhelpful: their roles overlapped with search, or Claude 3.7 Sonnet did the work itself rather than invoking them. After obtaining Claude Sonnet 4, the team tried again. Ball reported that it readily delegated clearly defined tasks, and the team added generic subagents that could write files and run terminal commands.[4]

Ball explains separate context windows as a practical advantage of delegation. A subagent can work through repeated attempts to fix a compiler error without filling the main agent's conversation with every intermediate step; multiple subagents can also work in parallel. Ball left the broader workflow consequences open. The team had deliberately avoided strongly nudging subagent use in its system prompt while exploring how best to integrate it.[4]

Ball also writes Register Spill, a weekly newsletter about software engineering and AI that began with timed Sunday writing sessions. He appeared alongside Quinn Slack in Latent Space's September 25, 2025 episode, Amp: The Emperor Has No Clothes, discussing Amp in its Sourcegraph context.[6][1]

Before his Amp work, Ball spent five years at Sourcegraph, worked at Zed, and returned to Sourcegraph. His earlier books provide technical context for his later implementation-focused teaching: both build runnable systems from the ground up.[2][5][3]

Ball wrote and self-published Writing An Interpreter In Go and Writing A Compiler In Go, dated November 2016 and July 2018 respectively. The first builds the Monkey programming language's lexer, parser, abstract syntax tree, object system, and evaluator, using tests first and no third-party libraries. The sequel continues with the same codebase, replacing the tree-walking evaluator with a bytecode compiler and virtual machine.[5]