TENSORLAKE·ENGINEERING LOG / FULL ARCHIVE
The full archive.
Every post we’ve published — engineering notes, benchmarks, model comparisons, and announcements. The curated feed lives on the main log.
Archive —CURRENT LOG
55 POSTS · NEWEST FIRST
Jul 22, 2026Engineering
Run Devin Outposts on Tensorlake Sandboxes.Cognition keeps the agent loop; execution runs inside a Tensorlake sandbox you control. We built the integration, and you can clone and run it today.
7 minJul 20, 2026Community buildsCrabbox Adds Tensorlake as a Delegated-Run ProviderOpenClaw's Crabbox — warm a box, sync the diff, run the suite — drives Firecracker microVMs through the tensorlake CLI, so one flag puts your test run in a sandbox.
2 minJul 13, 2026EngineeringFirecracker disk snapshots in O(changed bytes), not O(disk size)We added copy-on-write image layering inside Firecracker's block device, and the disk got faster, not slower. Snapshots pause the VM for tens of milliseconds and cost only what changed, so a running Postgres barely notices them, at any disk size.
10 minJul 11, 2026Community buildsBuilding Stateful AI Agents That Survive Session KillsAgent amnesia, solved: 84ms cold starts, snapshots as memory that carries prior experience, and forked sandboxes to evaluate candidate solutions in parallel.
10 minJul 10, 2026Community buildsYour Agent Benchmark Is Lying to YouAn 87% reported pass rate fell to 33% once verified — 53% of passes were fraudulent. Fork-per-task microVM isolation plus hidden verifiers make evals trustworthy.
12 minJul 10, 2026Community buildsWhy Your AI Experiments Keep Starting From Scratch (And How Tensorlake Fixes It)Warm up the environment once, memory-snapshot it, then fork N workers that resume mid-execution — instead of paying the same 40s setup tax on every run.
12 minJul 9, 2026Community buildsI Ran 5 AI Agents in Parallel on Tensorlake. The Isolation Held. Here Is How I Built It.Five agents in parallel, one deliberately crashed mid-run — the others finished clean with no defensive code, just per-agent microVMs. Registered images cut boot from 19.49s to 4.51s.
12 minJul 9, 2026EngineeringZero-copy TLS ingress with kTLS and splice(2) for sandboxes
8 minJul 6, 2026Community buildsBuilding Long-Running Claude-Managed Agents: Why State Matters More Than ComputeAn orchestrator/worker architecture on sandboxes that suspend when idle — checkpoint state and stop paying for compute instead of keeping boxes alive.
11 minJun 24, 2026EngineeringAccelerating Harbor with Tensorlake.Tensorlake runs the real task Docker images for Harbor on microVM sandboxes. On Terminal-Bench 2.1 the oracle passes all 89 tasks in 36m 44s — every image published, cold starts in seconds.
6 minJun 17, 2026Community buildsDemo: A Sandbox Agent Built on TensorlakeAn AI coding assistant that writes, runs, and exports code inside a Tensorlake sandbox — demo on X, code on GitHub.
1 minJun 17, 2026Community buildsSnapshot Once, Rollout a Thousand Times: A Practical RL Setup for Coding AgentsSnapshot the environment once and fork every rollout from it: 8,000 RL rollouts drop from ~20 hours of naive environment rebuilds to ~4.4 hours.
11 minJun 16, 2026Community buildsTensorlake Is Now a Sandbox Provider in DeputiesDeputies, an open-source control plane for background agents, added Tensorlake as a sandbox provider — after it met every criterion from the project's own "what makes a good sandbox" post.
1 minJun 16, 2026Community buildsWhy Most Multi-Agent AI Systems Waste 90% of Their Time (And How to Fix It)The bottleneck isn't concurrency — it's setup overhead. At 50 agents that's 4,500s spent preparing to work vs 50s with memory snapshots.
15 minJun 16, 2026Community buildsBuilding a Stateful Code Interpreter with Tensorlake SandboxesA build-along tutorial: from a toy code interpreter to one that remembers across sessions, suspends, snapshots, and forks — all on the free tier.
9 minJun 11, 2026Community buildsYour AI Agent Is One rm -rf Away From Disaster. Here Is What I Found After 5 Sandbox Experiments.Five isolation experiments running hostile, LLM-generated code. All four isolation boundaries held, and boots averaged 1.121s with only 148ms of variance.
8 minJun 5, 2026EngineeringTensorlake Sandboxes as the runtime for Claude Managed Agents.Run Claude Managed Agents on Tensorlake Sandboxes: you keep the control of self-hosting, and drop the always-on box. The short version — the server goes away.
7 minJun 4, 2026Community builds8 AI Agent Sandboxes, Ranked — and Why the Fastest One Will Wreck Your Agent in ProductionEight sandboxes benchmarked head-to-head. Tensorlake places fourth on raw cold-start speed but first on consistency: a 0.09s median-to-P99 spread vs 0.37s+ elsewhere.
11 minMay 26, 2026Community buildsI Built a Stateful Research Agent Inside a Sandbox. Here's What the Numbers Actually Looked Like.A research agent that keeps its state in a Tensorlake sandbox: it survives restarts and resumes mid-task, with real latency numbers throughout.
9 minMay 18, 2026ResearchClaude Opus 4.7 vs. Kimi K2.6: Real-World Coding TestA practical comparison between the premium closed frontier model and the cheaper open-weight agentic coder, on the same two-part coding task.
12 minMay 13, 2026EngineeringFilter and rank: how we schedule sandboxes across every cloudOne primitive — filter, then rank — runs through statefulness, multi-cloud, BYOC, and the data path. A tour of the scheduler behind Tensorlake's sandbox fleet.
14 minMay 1, 2026EngineeringBrowser Harness: direct Chrome DevTools Protocol access and a self-healing harness for browser agentsBrowser-use strips their harness to 592 lines and gives the LLM a raw WebSocket to Chrome's DevTools Protocol. When the agent hits a gap, it writes a helper function — and saves it for future runs.
3 minApr 30, 2026EngineeringHow scaffold design affects coding agent benchmark scores: lessons from DroidContextual instruction injection, per-model tool schemas, and planning/execution splits — the harness decisions behind Droid's 77.3% on Terminal-Bench.
3 minApr 29, 2026EngineeringHermes: the coding agent that gets better the more you use itMost coding agents are stateless. Hermes uses a closed learning loop where the agent creates skills from experience, improves them during use, and builds a persistent model of who you are across sessions.
3 minApr 29, 2026EngineeringPi: a coding agent with efficient system promptingPi keeps its entire system prompt — including all tool definitions — under 1,000 tokens, a 10× reduction over tools like Claude Code or Cline.
2 minApr 28, 2026EngineeringForgeCode: top open source coding agent in Terminal-Bench@2.0ForgeCode reaches 81.8% on Terminal-Bench 2.0 with both Claude Opus 4.6 and GPT-5.4 — a look at what the harness is doing that the model isn't.
3 minApr 27, 2026EngineeringBuilding Sandboxes for Computer UseComputer Use is a tiny loop. The hard part is building the boring, reproducible desktop around it.
8 minApr 24, 2026EngineeringStarting hundreds of sandboxes in parallel, and the design that makes it possible.We replaced reconciliation loops with durable command outbox to scale our sandbox scheduler to start 1000s of sandboxes every second
9 minApr 22, 2026EngineeringTensorlake is now an official Harbor environment runtimeHarbor defines and evaluates terminal tasks. Tensorlake provides the MicroVM execution layer. Together they are a full evaluation stack for CLI agents.
7 minApr 16, 2026ConceptsSuspend vs. snapshot: pause a sandbox, or save it for reuse?One is a pause button, the other is a save file. Same state, different question — and the answer shapes your cost model, your fan-out pattern, and which failures you can recover from.
12 minApr 14, 2026EngineeringHow we got to 5,000,000 sandboxes per projectA walk through the scheduler rewrite, the snapshot format, and the two syscalls that freed us from the noisy-neighbor tax.
12 minApr 09, 2026EngineeringSub-second cold starts for stateful microVMsSnapshot-restore beats boot every time. How we cut P50 wake-up to 80ms.
9 minApr 3, 2026EngineeringAutoresearch on steroids with sandboxesAn LLM agent can propose incremental training-script improvements, but safely executing untrusted code requires isolated sandboxes with resource limits — that's where Tensorlake comes in. Here's the overnight hill-climb, end to end.
9 minApr 02, 2026ProductRL environments are generally availableShip your reward model alongside the same sandbox that runs inference — no separate training harness.
4 minMar 27, 2026EngineeringDurable agent loops without a queueA checkpoint per step, at 2.4ms. Why we stopped writing to Redis and started writing to disk.
11 minMar 25, 2026EngineeringHarbor x TensorLake: Infrastructure for Agentic EvalsTensorLake is now integrated as a first-class environment provider in Harbor, enabling scalable agent evaluation with secure ephemeral MicroVMs.
2 minMar 21, 2026ResearchFilesystem benchmarks: sandboxed vs. bare metalSeven workloads, four runtimes. Most sandboxes leave 40% of your disk throughput on the table.
18 minMar 08, 2026ProductBYOC: run Tensorlake in your own AWS accountSingle-tenant control plane, PrivateLink, and data residency — without giving up the managed experience.
6 minFeb 28, 2026CompanyWe raised our Series A32M to build the sandbox layer every agent needs. A note on why, whats next, and who were hiring.
3 minFeb 19, 2026EngineeringFive patterns for tool execution in production agentsEphemeral vs. persistent, per-tenant vs. shared, and when to pay the cold-start tax on purpose.
8 minFeb 11, 2026ResearchPaged KV caches across sandbox boundariesA small experiment in sharing attention caches between sibling sandboxes without breaking isolation.
22 minFeb 9, 2026ResearchClaude Opus 4.6 vs GPT 5.3 CodexGPT 5.3 Codex excels at fast execution and agentic coding, while Claude Opus 4.6 stands out in reasoning-heavy and long-context tasks.
6 minFeb 04, 2026CompanySOC 2 Type II, and a primer on our security modelWhat we attest to, what we dont, and the exact kernel boundaries youre trusting.
5 minFeb 2, 2026ResearchOpenCode: The Best Claude Code AlternativeMost developers now use AI agents for coding tasks, but Claude Code's managed approach and constraints around model selection, pricing, and customization have spurred interest in open-source alternatives like OpenCode, which offers greater flexibility and control.
9 minJan 29, 2026EngineeringThe End of Database-Backed Workflow Engines: Building GraphRAG on Object StorageGraphRAG ingestion needs fan-out parallelism and durable retries that database-backed workflow engines can't give it; object storage can.
8 minJan 28, 2026ProductPython 3.12 is now the default imagePlus a new base image with uv pre-cached for 6x faster dependency installs.
2 minJan 19, 2026EngineeringBuilding HackerNews Podcast Generator with Gemini 3, ElevenlabsThis article shows how to build a simple podcast generator that turns Hacker News posts into short audio summaries using a single Tensorlake Application.
8 minJan 18, 2026ResearchOpenAI GPT-5.2-Codex (high) vs. Claude Opus 4.5 vs. Gemini 3 Pro: Real-World Coding comparisonThree leading AI models were tested on real coding tasks in an 8K+ star repository to compare code quality, reliability, and practical performance.
10 minJan 14, 2026ResearchClaude Cowork: Architecture, Capabilities, and Usage OverviewClaude Cowork is an agentic task execution mode allowing Claude to plan and complete multi-step work on your behalf.
8 minJan 8, 2026EngineeringBuilding an Outage Agent That Handles Repeated Production IncidentsAn outage agent built on Tensorlake Applications triages alerts, investigates with Groq and Exa, and only escalates when a human is actually needed.
9 minDec 28, 2025ResearchThe Next Enterprise Platform Isn't Data-Driven, It's Context-DrivenContext graphs capture the reasoning behind decisions, not just their outputs, making AI-driven workflows auditable without replacing existing systems of record.
11 minDec 15, 2025ResearchGemini 3 vs GPT-5.2: Detailed Coding ComparisonGemini 3 Pro is fast and multimodal but produces simpler output; GPT-5.2 trades speed for more reliable, production-ready code.
8 minDec 13, 2025ResearchI Tried Letting Antigravity Build An Agent For Me. Here's What Actually HappenedWe let Antigravity build a guest checkout flow and abandoned-cart recovery emails end to end, and tracked where it held up and where it didn't.
8 minDec 9, 2025ResearchOpenAI GPT-5.2 Codex vs. Gemini 3 Pro vs Opus 4.5: Coding comparisonGPT-5.2 Codex, Gemini 3 Pro, and Claude Opus 4.5 build a Minecraft clone, a Figma dashboard, and solve a hard LeetCode problem, with very different results.
8 minDec 9, 2025EngineeringTOON vs JSON: A Token-Optimized Data Format for Reducing LLM CostsTOON strips the repetitive syntax out of JSON for LLM prompts, cutting token usage while improving structured extraction accuracy.
8 minLegacy —DOCUMENT AI ERA
19 POSTS
Posts from before Tensorlake became the sandbox cloud for AI agents, kept online for reference. For what we build today, start with the current log.
Jan 20, 2026Engineering