novosad.xyz

whoami

Bogdan Novosad. Staff Software Engineer/Agentic development/Go, Rust, C#/.NET

I help engineering teams move to agentic development: agents write the code, engineers decide what to build and check that it is right.

I did this at my company, on a learning management platform used by schools across Europe: I designed the shared multi-agent workflow, write all my own code through it and teach it to other engineers in seminars. With the same approach I build and ship my own products in Go and Rust.

I have 13 years of experience in backend systems, SaaS products and developer tooling. Open to Senior/Staff roles, full-time or B2B contract. Remote in Europe, or on site after relocating to Madrid or Barcelona.

~/workflowHow I ship with agents

One agent keeps the plan and calls a specialist only when it pays. Every file has one writer, a fresh reviewer checks the result, and a person decides what ships.

  1. [1]

    Scope

    One task per session. The agent keeps a short file with the criteria, the decisions and what is left, so a new session can read it and continue.

  2. [2]

    Plan

    The main agent plans and keeps the decisions. Before a risky stage it writes down which test or check will prove each requirement. A hard question goes to an advisor.

  3. [3]

    Write

    Small edits stay with the main agent. Otherwise one writer per set of files, in parallel only on parts that do not overlap.

  4. [4]

    Review and verify

    An agent that has not seen the reasoning reviews the final diff. Long test runs and browser checks go to a cheaper model.

  5. [5]

    Ship

    The main agent commits, pushes and opens the pull request itself. It merges, releases or deploys only when told to.

Claude Code · 4 roles on four model tiers · token use measured per role · measured here · open source

~/projectsProjects

Wearbrands

creator · 2026

A fashion search across shops: type an ordinary query such as "black dress" and get clothes from several shops in one place.

  • Go API and ingest pipeline, PostgreSQL with pgvector, a SvelteKit front end, Terraform on AWS.
  • Archive before parse: every scraped response is saved first, so a parser fix is tested by replaying saved responses. A new shop must pass a small probe run before a full run, and tests never call the paid API.
  • Go
  • PostgreSQL
  • SvelteKit
  • AWS
wearbrands.co →

TeoriaYa

founder · 2026

Practice for the Spanish driving-license theory test, for people who are still learning Spanish: learners answer the questions in Spanish and can see each one translated into their own language. Built end to end with AI agents.

  • Go API as a modular monolith with separate HTTP and AI-worker roles; PostgreSQL with sqlc and golang-migrate.
  • Content pipeline with AI-assisted rewriting, images, translations and glossary tooltips, each behind a human review gate.
  • Rulebook grounding: retrieval over the official traffic rulebook, with local embeddings and an evaluation set, finds the rule behind a question, so content is checked against the source.
  • Go
  • PostgreSQL
  • SvelteKit
  • RAG
teoriaya.com →

GliderMCP

creator · 2026

Local MCP servers that let AI coding agents ask a codebase exact questions: where a symbol is defined, who calls it, and what a change would break.

  • Glider (C#/.NET): Roslyn-powered navigation, references, call graphs, diagnostics and refactoring. About 148K NuGet downloads.
  • Scout (Rust): local code index with literal, regex and fuzzy search, structural search across nine languages, and optional on-device semantic search with GPU acceleration.
  • Keep (Rust): shared knowledge store on SQLite with role-based access and an audit log.
  • Rust
  • C#
  • TypeScript
  • MCP
glidermcp.com →

~/writingWriting

ls -lt writing/

  1. Nov 2026

    Why not one agent on the best model? When a subagent is worth it

    A short guide to one decision: do the work in the main agent or hand it to a subagent. Four common reasons for delegating, checked against what Anthropic, OpenAI and Google publish and against my own sessions.

  2. Nov 2026

    Where the tokens went: rules of thumb for an agentic workflow

    I measured my own agent sessions. Half of the main agent's cost was re-reading its own history, a quarter of one session was advisor calls I had not counted, and eight independent reviews were 6%. Five rules follow.

all writing → · feed.xml

~/experienceExperience

git log --career

  1. 2024 — now

    Staff Software Engineer · Software Country

    Learning management platform used by schools and universities across Europe (~7M daily active users). Leads agentic development in the company.

  2. 2018 — 2024

    Senior Software Engineer · Aspose

    Reimplemented Microsoft Word layout algorithms in Aspose.Words; designed and built an online document viewer SaaS with .NET, Angular and PostgreSQL.

  3. 2013 — 2018

    Full-Stack Engineer · ADICOM

    ASP.NET MVC monolith and new .NET Core microservices with Vue.js frontends.

~/contactGet in touch

bogdan@novosad.xyz