eliaz : 2026 · living preprint · v3
OCT 2026
Yossi Eliaz
38live pages here
15years on GitHub
✦ Selected work
Now building
Systems for coding agents, and one genome-scale measurement. Each states its limits.
THESIS 001 • 2026
THE LIVING LAYER
A speculative architecture thesis.
What if sandboxes were the universal execution primitive, and one governed database held UI, logic, state, history, and policy?
Software would be instantiated, not deployed.
What if sandboxes were the universal execution primitive, and one governed database held UI, logic, state, history, and policy?
Software would be instantiated, not deployed.
✦ Featured essay · Jun 2026
Databases in the Age of AI
What fifty years of database systems must become when the dominant reader of data is no longer human.
Read it live →
Fig. 00 — captured live
§ 01
AI Agents & Sandboxes
Field guides, demos, and measurements on agent execution. Every figure is a live page.
Fig. 01● live pageThe Sandbox ShiftWhen isolated execution helps agent systems, and how to choose an approach, with an interactive decider.View live →
Fig. 02● live pageCI Caching Is Not One CacheA June 2026 POC: when native caches, BuildKit, or Incredibuild Build Cache help GitHub Actions on fresh, warm, or touched runners, with POC timings (not a controlled benchmark).View live →
Fig. 03● live pageWasted CyclesFind the machine time blocking agent work: local traces and GitHub Actions latency, with a one-line installer.View live →
Fig. 04● live pageHermetic AgentsDeterministic demo: scripted code and test writers get the same spec but disjoint inputs; a QA arbiter with a hidden oracle blames the seeded wrong test. No language model is called.View live →
Fig. 05● live pageBazel 9 × montyBuilding pydantic/monty with Bazel 9 and a remote cache: 121 s cold, 15.7 s fully cached.View live →
Fig. 06● live pageAI Sandbox LandscapeA map of the execution-sandbox ecosystem.View live →
Fig. 07● live pageOne Lab Per KernelA technical design for running a GPU-kernel research loop (SOL-ExecBench) reproducibly on Islo sandboxes.View live →
Fig. 08● live pageRust It UpPrototype: paste a public GitHub repo to get an illustrative Rust-migration scorecard. The scanner is mocked; estimates are not audited.View live →
Fig. 09● live pageInferoa Flight Recorder (unofficial demo)Unofficial demo of agentic-in/inferoa: a client-side simulation of prefix caching, context compression, and routing, plus receipts from real vLLM runs in islo.dev sandboxes.View live →
Fig. 10● live pageAgentic AirbyteDesign for agent-driven ETL, plus one verified run: 50,000 rows from ClickHouse to DuckDB in an islo.dev sandbox. Not affiliated with Airbyte.View live →
Fig. 11● live pageEvals or You're BlindRunnable demo after Joshua Görner's PyCon DE 2026 talk: Pydantic Evals grade a Pydantic AI build-triage agent offline, live, or on a disposable crabbox box.View live →
Fig. 12● live pageRepo2RLEnv × CrabboxDemo record: scoring Repo2RLEnv pr_diff tasks on an islo.dev sandbox via crabbox, about 50 s per task.View live →
Fig. 13● live pageOdysseus × CrabboxDemo: run odysseus-dev's Odysseus AI workspace, and its agent's code, on throwaway islo.dev microVMs via openclaw/crabbox, from a fork's one-file crabbox.sh.View live →
Fig. 14● live pageTiny Long-Memory Benchmark (toy demo)A LongMemEval-style toy benchmark fanned out across sandboxes.View live →
Fig. 15● live pageMeta-harness on IsloA deterministic simulation of harness optimization: scripted agent, pattern-matching proposer, five toy tasks.View live →
Fig. 16● live pageenv-rosettaDemo: @adithya-s-k's Wordle env from RL_Envs_101 in four RL frameworks (OpenEnv, ORS, NeMo Gym, Verifiers); two ran in islo.dev sandboxes, and the sample rollouts are scripted.View live →
Fig. 17● live pageCrabbox Runner for RustRoverMVP plugin: run cargo test and clippy from RustRover through the OpenClaw Crabbox CLI on remote Crabbox runners or Islo sandboxes.View live →
Fig. 18● live pageGemma System OneResearch prototype: Gemma 3 4B as a one-forward-pass decision function in native Rust, with shared-state inference across many questions.View live →
Fig. 19● live pagePokeloopDemo: a genetic-algorithm loop over LLM system prompts, designed to fan out across parallel islo.dev sandboxes, with a Pokémon Crystal-style viewer. Results are a scripted playback; no agent was trained.View live →
Fig. 20● live pageunity-loopProof of concept: Claude proposes and scores CSS variants around one fixed Unity WebGL scene (Lightning VFX by @MirzaBeig), each served from its own islo.dev sandbox. Scores shown come from a deterministic offline demo.View live →
Fig. 21● live pagewolfram-fb0Can a 1B local model write x86-64 assembly for Rule 30 inside a VM? First 3 iterations: not yet.View live →
§ 02
Science & Genomics
Computational genomics, open physics, and evidence-aware compute.
Fig. 22● live pageReference Broadcast by VM SnapshotDemo on islo.dev: warm one sandbox with the GRCh38 reference, snapshot it, then fork it copy-on-write per chromosome for genomic scatter-gather.View live →
Fig. 23● live pageOpen Physics Problems112 open physics problems mapped like open mathematics: publication cultures, refutability, and ten-day agentic sprints.View live →
Fig. 24● live pageEvidence-Aware MapReduceFork workers freely while keeping evidence, uncertainty, and shared origins visible, so a repeated error is not counted as consensus. Preprint, 2026.View live →
§ 03
Deep Dives & Essays
Longer reads and field reports on how software is being rearranged.
Fig. 25● live pageThe Stochastic CPUA field report on the cloud being rebuilt for agents: the agent hour, two-layer caches, snapshot provenance.View live →
Fig. 26● live pageFrom Microservices to MicrocellsA plain explanation of microcells: tasks that run when needed, save progress, and survive a crash, plus a safe migration path.View live →
Fig. 27● live pageFree* IntelligenceA field report on the difference between a fluent model answer and an accountable agent harness.View live →
Fig. 28● live pageThe Living LayerA speculative architecture thesis: sandboxed execution plus governed, queryable state.View live →
Fig. 29● live pageThe Data That Taught the MachinesHow Stack Overflow trained its own replacement.View live →
Fig. 30● live pageThe Loop Is Coming Apartjj, Crabbox, and the unbundled development cycle: the diff is the unit of work.View live →
Fig. 31● live pageIntel: Vision Without ExecutionAn opinion essay in comic style on Intel's bets and misses, 2000–2026.View live →
§ 04
Talks
Slides and talk pages.
Fig. 32● slides & codeYour Agent Escaped Without Escaping the SandboxFour doors an agent can walk out through without breaking the wall: the key, the folder, the wire, the verdict. Zenity AI Agent Security Summit, New York, October 2026.View on GitHub →
Fig. 33● slides & codeForkable Sandboxes: The Runtime Layer for AI Software FactoriesSeminar at the Cambridge Computer Laboratory Systems Research Group, 15 October 2026: forkable, isolated execution for coding agents. Slides, scripts and references.View on GitHub →
Fig. 34● live pageA Sandbox Provider for Apache AirflowOne ephemeral microVM per Airflow task try, for fast fan-out and controlled untrusted execution. Airflow Town Hall, July 2026.View live →
Fig. 35● live pageA Million Compiles. One Robot Hour.Rust China Conf talk page in English (中文 version linked): a simulated robot, a stale-observation bug, the Rust fix, and the Incredibuild rebuild loop, with recorded evidence.View live →
§ 05
Quant Finance
Market microstructure, made touchable.
Fig. 36● live pageHow Market Making Works13 in-browser simulations, from the limit order book to Avellaneda–Stoikov, Glosten–Milgrom, and Kyle's λ.View live →
Fig. 37● live pageAvellaneda–Stoikov, in a SandboxA synthetic reproduction of the Avellaneda–Stoikov vs symmetric market-maker comparison from Sasson, Ho & Samson (Stanford MSE448, 2018), run in an Islo sandbox in about 90 seconds with no local Python. The qualitative result reproduces; the dollar magnitudes do not.View live →
§ 06
Receipts
Public activity, continuously updated.