~/jiri-studnicka $ whoami

Agentic Engineer

I'm Jiří Studnička. I spent twenty years building backends for banks, insurers and energy companies. Now I build systems where AI agents write, test and ship production software — and I vouch that it holds.

// building agentic systems since GPT-3.5 — before it was cool

// depth −7 m · sediments of practice

You draw better from the deep

Studnička is Czech for a little well. It took twenty years of enterprise strata to dig — and that is exactly the depth every agent I set loose on production code draws from.

  1. 2004

    First lines in production

    IIS Tábor — Gupta and .NET. The craft begins.

  2. 2008–15

    The .NET era & a first company

    Capgemini, Minerva, RADIUM. Co-owner and team leader at 4Leaders.

  3. 2015–18

    Leveris — banking startup

    Scala, Akka, Play, Kafka, MongoDB. Functional programming, monads, event sourcing. A bank built from scratch.

  4. 2018–19

    E.ON SolarCloud

    Scala and Akka Persistence for the energy sector — event sourcing in live production.

  5. 2019–20

    Home Credit — big data

    Spark, Kafka, HBase, Hadoop, Hive. Integrating the primary systems and DWH of a multinational group.

  6. 2020–22

    Generali, SOFTINO

    Java, .NET, Python. Spring Boot, Oracle, Azure Synapse, Spark.

  7. ~2023

    GPT-3.5 — the break

    I understood the craft had just changed. I've been doing agentic engineering full-time ever since. You don't throw seniority away — you reforge it.

“Most AI engineers discovered Claude Code yesterday. Meanwhile I'm working out how to keep a fleet of agents running on a production repo for three days straight.”

GPT-3.5ClineRoo CodeAugment CodeClaude CodeCodexOpenCodecustom orchestration…and a pile of those nobody remembers anymore

// depth −18 m · the tool seam

The Smithy

I named my tools after the smithy. Not for the romance — because agentic engineering is a craft: it has a process, tools and quality control. This is my production cycle.

01 · the anvil

Anvil

issue → spec

Hammers a raw GitHub issue into an implementation-ready spec. Agents read the assignment against real code, hunting gaps and ambiguities — and a human approves every write.

  • read-only analysis
  • human-in-the-loop

02 · the forge

Forge

issue → pull request

An orchestration platform: agents run through spec, implementation, code review, validation and CI loops — all the way to merge and the checks beyond it. No “trust us”: every requirement has its ledger.

  • 27-state state machine
  • requirement ledger
  • 50+ CLI commands

03 · the quench

Temper

PR → proof

Self-healing testing in a real browser. When a selector drifts, the agent understands the original intent, the test completes — and the scenario repairs itself to run deterministically next time.

  • screenshot of every step
  • self-healing scenarios

CogneraSpec

A spec-driven CLI that holds agents to their word: every requirement carries a stable REQ-ID and verifiable evidence in tests. Drift between spec and reality is measured, not guessed.

AI Toolkit

A library of 59 commands and 25 skills for Claude Code and Codex — paired audits and fixes for security, performance, a11y and more. Standardised agentic workflows.

Meeting Agent

Meeting recording → ready-made GitHub issues. Gemini reads the video including the screen, Claude writes issues in the repository's context, screenshots attach themselves.

…and more: agent-system (orchestration via Telegram), MCP servers, invoiceAI.

01

Spec-first. An agent without a proper assignment is a random code generator.

02

Proof, not vibes. Every requirement has a test, evidence and a screenshot.

03

Human-in-the-loop. AI writes. A senior vouches.

// depth −29 m · load-bearing masonry

Where it all stands

// depth −42 m · the waterline

At the bottom of every well there is water

Looking for someone who actually runs agentic development — not someone who just posts about it on LinkedIn? Lower the bucket.