Michal
Hornicky

Senior Software Engineer /Agent Infrastructure /ML Systems
Michal Hornicky
00 / Current Focus

The question I'm working on: What is the optimal topology/harness/meta-harness that enables Agents to recursively and reflexively improve themselves? This includes creating better RL environments, model architectures, algorithmic improvements or better harnesses.

01 / About
  • Engineer at Google DeepMind on the Gemini team; ten years across ML systems, distributed infrastructure, and large-scale applications
  • I work prototype-first: try the idea, survey the state of the art, then turn the exploration into a design doc and a shipped system
  • I optimize for automatability: every other metric can be improved if the process can be automated.
02 / Experience
Senior Software Engineer · Gemini / Google DeepMind
Nov 2024 – Present
Google · Zürich
  • Technical lead of a small team working on Self improving agent infrastructure for the Gemini application: agent development environments where fleets of collaborating agents perform SWE tasks, and productively improve themselves
  • Early adopter of agents across the Gemini application stack; built many of the early integration systems: multi-agent orchestration, agent harnesses, RL environments, specialized sandboxes
  • Drive adoption through tooling and workshops; the job is researcher and engineer velocity, making the people around me faster with better machinery
  • Now focused on enabling self-improving agentic infrastructure
Software Engineer · Assistant & Gemini
Nov 2021 – Nov 2024
Google · Zürich
  • Developed automated test infrastructure and tooling for Android and iOS applications
  • Supported testing and releases for Google Assistant, Search, and Gemini/Bard mobile apps
  • Built early internal LLM tools for automated testing and software verification
Senior Software Engineer / Team Lead
Jul 2020 – Sep 2021
Red Hat · Brno
  • Led a team of 5 engineers building the Patch application on Red Hat Insights (cloud.redhat.com)
  • Engineered a custom in-memory dependency solver and evaluation engine for RHEL package resolution after PostgreSQL hit its limits on graph traversal
  • Scaled the platform to serve thousands of RHEL fleet deployments daily
Software Engineer
Sep 2019 – Jul 2020
Red Hat · Brno
  • Initiated development of the Patch application; built Go libraries for integrating into the Python ecosystem
Android Developer
Mar 2016 – Sep 2019
TouchArt s.r.o · Brno
  • Full lifecycle management of 10+ B2B Android applications; led architectural modernization to Kotlin, MVVM, reactive dataflow
03 / AI Projects
Bedrock · Autonomous Software Engineering Platform
semtexzv.github.io/bedrock
A fleet of coding agents that builds, reviews, and ships a software platform end to end · 2026 – Present
  • A self modifying control plane for coding agents: a fleet of ≈15 collaborating agents that build, review, and integrate changes into a shared monorepo, end to end
  • Built on Buck2 and jj: a self-integrating monorepo with hermetic builds, CI gating, safety mechanisms
  • Generates its own data: every change, dispatch, conversation, and review is a logged trajectory. The direction is the platform as an RL environment, with rubric-scored trajectories (rubrics as rewards) and continuous post-training of its own agents on its own data
  • Bounded self-modification, designed against reward hacking: a structural security layer protects the platform itself
llm-lab · Exploratory LLM Training
github.com/semtexzv/llm-lab
A home-run LLM architecture research platform, executing an autoresearch loop targeting small models · 2025 – Present
  • Verifies or falsifies my hypotheses about LLM architecture, with the goal of finding more efficient architectures that reduce training costs; results are analyzed on FLOPs, memory, bandwidth, and token efficiency at 1B scale
  • Currently focused on novel architectures of the residual stream in decoder-only transformers, inspired by manifold hyper-connections (arXiv:2512.24880) and attention residuals (arXiv:2603.15031)
mage · Self-Recompiling Coding Agent
github.com/semtexzv/mage
An agent that extends itself by recompiling its own binary
  • A Darwin Gödel Machine-style self-modifying agent at the compiled-artifact level: embeds a snapshot of its own source, splices in new modules, recompiles a fresh binary, and hot-swaps into the next generation
  • Project intended to investigate self-modification, informed much of my further RSI work.
Termineer · Autonomous AI Coding Agent
github.com/semtexzv/termineer
Self-directed coding agent built on a configurable tool-calling grammar
  • A provider-agnostic function-calling abstraction: one configurable grammar unified tool calling across Anthropic, Google, and OpenAI, built when providers had no shared format for it
  • Autonomous task decomposition with specialized sub-agents; MCP protocol support
GenAgent · Always-On Personal Assistant
github.com/semtexzv/genagent
Taskbar AI assistant for macOS/iOS, explored alongside Termineer
  • Menu-bar assistant invokable anytime for coding and general help
04 / Distributed Computing Infrastructure
Rube · Generalized Control-Loop Architecture
github.com/semtexzv/rube
  • A generalized reconciliation and control-loop runtime: the Kubernetes operator pattern applied to arbitrary resources, desired-state management without the YAML or the containers
Protokit · High-Performance Protocol Buffers for Rust
github.com/semtexzv/protokit
  • Only Rust implementation supporting the full TextFormat specification; competitive with prost; dual front-end compiler
Stor · Embedded Key-Value Database
github.com/semtexzv/stor
  • Typed key-value database on RocksDB with zero-copy deserialization; storage engine for a low-latency portfolio platform
05 / Skills
Agent Systems
Autonomous coding agentsMulti-agent orchestrationAgent harness designRL environmentsContext engineeringTool callingMCP protocol
ML / AI
PyTorchMLXPost-training (SFT, RLVR, GRPO)EvalsTraining pipelinesDistributed trainingReinforcement learningTransformer architecturesAutodiff
Systems
Distributed systemsHPCGPGPU (CUDA / OpenCL)ML acceleratorsHermetic buildsSandboxingActor models
Infrastructure
KubernetesTerraformLarge-scale data processingObservabilityCI / CDGCP
Languages
PythonRustGoC / C++SwiftKotlinTypeScript
Spoken
Slovak (Native)Czech (Native)English (Fluent)German (Beginner)
06 / Education
Master's Degree, Information Technology
2017 – 2019
Brno University of Technology · Brno
  • Thesis: distributed algorithmic trading system (Rust, ZeroMQ, NATS, Kubernetes)
Bachelor's Degree, Information Technology
2014 – 2017
Brno University of Technology · Brno