DigitalOcean

Your agent can write the code. Can it actually open the pull request? 👀 The second an agent starts acting, every tool it touches needs a credential somewhere. DigitalOcean Action Gateway gives agents governed access to external tools without exposing the credentials. It doesn't care what harness you run. Anything that speaks MCP hits the same endpoint. No tool catalog gets loaded into your model's context. The token itself never reaches the model or the sandbox. And when something fails, you get a real error back: class: unauthorized, retriable: true, recovery_hint: refresh_auth. We tested it end to end: an agent cloned a repo, cut a branch, committed, pushed, opened a PR, and merged it, all without touching the actual credentials. Harness Runtime hosts the agent and Action Gateway governs what it's allowed to touch. Check out the full breakdown. 🔗 https://do.co/4haLRh9

Can a training run survive its GPU being taken away mid-run? Yes. Zero steps redone. ✅ Developer Advocate, Narasimha Badrinath, fired a reclaim notice mid-run on a full-parameter Qwen3-8B fine-tune, running on an AMD Instinct™ MI350X spot GPU Droplet. The trainer finished its step, the 98.3 GB checkpoint hit DigitalOcean Spaces 5.4 minutes later, and a replacement GPU picked up at step 14,625. It worked because nothing needed to resume lived on the GPU itself: weights, optimizer state, scheduler, data cursor, RNG all left the box. Full timings, costs, and what broke along the way.🔗 https://do.co/4iMXrQQ

pgvector looks fast on a few thousand test rows, but that's not the number to plan around. Latency stays flat, then jumps hard, usually not at some magic row count but when your HNSW index outgrows memory and Postgres starts reading it off disk. Because it's a cliff and not a slope, teams don't notice until it's already hitting production. A few things worth knowing before you hit that wall: • Index build time usually breaks before query time. Continuous re-indexing means you pay that cost on every rebuild. • IVFFlat doesn't degrade evenly. Real embeddings cluster unevenly, so you get random spikes instead of a steady decline. • Approximate indexes trade consistency for speed. The same query can return different rows across runs. • ef_search and probes trade off speed against recall. You can tune for one, not both. We break down how to test for your specific breaking point before it finds you first. 🔗 https://do.co/4ykI1YI

Vercel and DigitalOcean are teaming up to bring GLM-5.3 to AI Gateway users at half the cost, for one week only. Offer details available on Vercel AI Gateway. https://do.co/4qPma98

Your clipboard just got smarter. 👀 Haimantika Mitra, Senior Developer Advocate, built a tool that watches whatever you copy and acts on it automatically. Copy a link, it summarizes the page. No app to open. No prompt to write. The tool classifies your clipboard content and responds accordingly. It uses a Nous Research Hermes agent running DeepSeek AI-V4-Flash on DigitalOcean Serverless Inference. No model to host, no GPU to manage.

Something breaks in production at 2am. The first twenty minutes are always the same: read the traceback, find the line, write a test, fix it. That is not a judgement call. It is a checklist. Our developer advocate, Amit Jotwani, handed that checklist off. When his app crashes, it posts the traceback to a trigger in DigitalOcean Managed Agents. The trigger boots a fresh sandbox, runs a coding agent against the error, and tears the sandbox down when the run finishes. It found the bug, wrote a test that fails without the fix, fixed it, and opened a pull request. He was making a coffee. ☕️ Triggers fire on a webhook or a schedule. Now in private preview. 🔗 https://do.co/4cokbmt

Claude Fable 5.1 is now available through the DigitalOcean Inference Engine, day zero. ⚡️ https://do.co/4yk6I7M It's Anthropic's most advanced model for coding and knowledge work, now with 75% cheaper cache-read pricing, making long-running agents that replay large context far more economical. Available via Serverless Inference, Inference Router, model synthesis, and Evaluations: fully managed, usage-based pricing, no separate Anthropic contract.

Your coding agent dies the moment you close your laptop. So we built DigitalOcean Managed Agents Runtime Services (M.A.R.S.). One doctl command drops Claude Code, Codex, OpenCode, or your own LangGraph agent into an isolated sandbox. Pause it, close everything, resume tomorrow. Workspace and full conversation intact. Checkpoints snapshot the machine and agent memory together. Fork one session into three, try three approaches on the same refactor, keep the winner. Action Gateway handles tools: GitHub, Jira, Stripe, 1,000+ more behind a single MCP endpoint. Your keys stay in Secrets Manager until a tool runs, never touching the model. Sign up for the private preview today. 🔗 https://do.co/4cokbmt

Our Solutions Architect Darian Wilkin put Z.ai's GLM-5.3 to the test in Flotilla, a naval resource-and-diplomacy simulation where four LLMs write and rewrite their own ship-control code across a five-game series, this time against Kimi, DeepSeek, and Qwen. A few standout moments from GLM-5.3's run: • Won the series 3-2, closing game five with its highest score of the series • The only model to use dispatch cutters, relaying new orders to ships already out at sea • Diagnosed and rewrote its own broken ship code mid-run rather than stalling • Negotiated, broke, and re-negotiated non-aggression pacts as the game state shifted, timing an offensive push to when it had the advantage Try it yourself: https://lnkd.in/gQgqkCQj

Curious about GLM-5.3-Flash? Our Senior Developer Advocate, Ryan O'Connor, walks through what makes this model from Z.ai worth a look for agentic workloads. A few highlights: • Scores 57 on the intelligence index, delivering top-tier agentic coding performance at a fraction of typical cost • First natively multimodal model in the GLM series, enabling browser and computer use so it can inspect UIs and critique code visually • 1M token context window, MoE architecture (320B total / 18B active parameters per token) • Nearly doubles GLM-5.2's performance on AutomationBench It runs fully managed on DigitalOcean's Inference Engine; swap one endpoint and start testing it against your own workloads.