The LLM Whisperer

Javier Mellado · Agentic AI engineer

I install AI agents in engineering teams, and the team keeps the capability.

21 years shipping software. Now I run fleets of coding agents from ticket to merged pull request, and I teach teams to do the same. Based in Edinburgh, UK.

[email protected] See the pipeline Read the method

Built in the open

From issue to merged pull request, with the humans at the gates

Agentic Team Member is my delivery pipeline. An agent takes an issue, starts from a failing test in an isolated clone, passes the checks, gets the code nobody asked for cut, and hands the branch to a review gate from another lab. A human owns the merge. La Roca is the memory underneath: everything the agents did, searchable by meaning.

A month in the factory

6,914agent sessions in a month
250pipeline runs in ten days across two codebases, 155 of them delivered to the review gate
86pieces of slop caught and cut before merge, recorded across all projects
111pull requests merged in Agentic Team Member in ten days
93pull requests merged in La Roca in a month
9different agent tools at work in the same month

Counted from La Roca and the repositories, September and October 2026.

How the pipeline works →

In teams

What usually happens when I join a team

Delivery from the first week

  • A first pull request merged within days of joining a codebase I had never seen.
  • Production code on the live product, with proofs of concept in parallel, such as natural language over the company's data, or plain English in and Playwright tests out.
  • Agentic development from day one: coding agents on real tickets, and every change through the team's own review.

The team's way of working changes

  • AI review becomes a normal step before every human review.
  • People who pair with me change how they work.

The Whisperer Method

Your first 30 days in a codebase you have never seen, with agents

Distilled from real starts. Seven habits any engineer can follow from day one, and six things the team prepares before they arrive.

  1. Day 1 is the environment, and it gets written down that day.
  2. Map before code: agents explore the codebase in parallel.
  3. A context file for agents from the first day.
  4. A small first ticket at the edge, to walk the whole path to merge.
  5. Ask the architecture question before touching anything.
  6. Three failed attempts: stop, revert, cut the task down.
  7. Week three: give the documentation back to the repository.
Read the method (PDF) →

Open source

The tools behind the work

agentic-team-memberThe delivery pipeline: from issue to pull request with an isolated clone, a failing test first, checks, cuts and an independent review gate.
la-rocaLocal memory for agent fleets: every session your agents left behind, searchable by meaning. One file, no dependencies.
slopslintA blocking slop gate for AI assisted repositories: duplication, tombstones and ceilings that only go down.
human.mdA framework for healthy collaboration between people and coding agents.

Work with me

Your product keeps moving while your team learns to work with agents

BuildFeatures and fixes on your live product from the first week, through your own review.
SupportIncidents, data questions and the daily work around a live platform.
ResearchA proof of concept of natural language over your data.

And after: a regular advisory session with your team on AI patterns, tools and roadmap. Or start with a four hour session, from spec to pull request on one of your own tickets. Write to [email protected].