Bernat Sampera
Skip to writing

Writing on AI.

How we built a context tree for our agent to resolve support tasks

Support work led us from manual fixes to a tree of context files. Each index points the agent to service details, task instructions, and past investigations.

Read the article
Follow the context tree to a support runbook The root index links to Stripe, Firestore, and Support. The Support index links to runbooks and logs. A runbook describes how to resolve a membership mismatch. llms.txt stripe/ firestore/ support/ runbooks/ logs/ fix-membership-mismatch.md

Express the General

State what you want and the constraints that matter. A short essay on prompting, inspired by Kierkegaard, and why guesses about the method can hide the goal.

Read the article

Your AI Agent Deserves Its Own Repo

Keep agent knowledge in version-controlled files. Organize the context, load what each task needs, and save repeated work as reusable procedures.

Read the article
Load useful context from a version-controlled repo The repo contains integration, workflow, and task modules. A session can load the relevant file. The knowledge stays in the repo for later sessions. Integration: Stripe setup Workflow: deploy steps Task: active bug Load what the task needs. The files stay. The context can evolve.

Lessons Learned Building a Real-World AI Agent with LangGraph

A translation agent needs room for feedback. This LangGraph workflow keeps the conversation in state and moves slow glossary suggestions into a background job.

Read the article
A translation loop with a separate background job An initial translation reaches the user. The supervisor waits for feedback. A refinement returns to the user. A separate background job suggests glossary changes. Translate Ask the user Refine The conversation continues. Suggest glossary changes

Advanced Langgraph: Deep dive into open deep research

An analysis of how Open Deep Research splits work across agents and carries context between steps. Includes architecture tradeoffs, limits, and practical takeaways.

Read the article

BJJGym

bjjgym.com

Find a BJJ gym. Wherever you travel.

Find Brazilian jiu-jitsu gyms around the world and choose where to train.

Problem
Gym listings often lack useful details about training styles, cleanliness, teachers, and training level.
Solution
Use AI to extract labels from public reviews, such as no-gi, clean, good teacher, and high level. Show these details for each gym.
“Dropped in for no-gi. Clean mats, a coach who explained every step, and tough rounds.”
No-giCleanGood teacherHigh level
Example review and labels.

IndoEuroMap

indoeuromap.com

Explore a city. See its local patterns.

A map of area profiles built from public place and review data.

Problem
Scattered place and review data are hard to understand at the neighbourhood level.
Solution
Combine public signals into area profiles and show them on a map. These profiles do not establish an individual’s ethnicity or culture.
Illustrative map.

gcontext

gcontext.ai

A new session. The knowledge stays.

Shared, structured context for AI agents across sessions.

Problem
Useful knowledge gets lost between sessions or stays scattered across tools.
Solution
Store knowledge, connections, and procedures in plain files that agents can load and update through MCP.

The task ends.
Keep the useful context.

01 Decisions saved02 Procedures reusable03 Connections available

Ready for the next session.

Example context structure.