Yaniv Markovski

I am excited about artificial intelligence, technological advancements, education, and customer experience. Ex-OpenAI, currently at AI21 Labs.

[Yaniv Markovski @yanivm13]

@nimble_search Go Nimble!

[Nimble @nimble_search]

Introducing Web Search Agents: expert level web search that self-learns your domain to deliver perfect web context, while burning less tokens.

Generic web search tools start from scratch every time they run. Web Search Agents compound their expertise to adapt web search and crawling to your use case — finding the exact data you need, more efficiently.

Learn more below in our blog: https://t.co/Vdy9sM1eSW

[Yaniv Markovski @yanivm13]

@pamelafox MCPJam looks super handy for iterating on tool calls and OAuth. next pain point after you get requests working is state: parallel runs, rollback, and safe writes. we wrote about workspace isolation for MCP tools here:
https://t.co/GSqp31UxVT

[Yaniv Markovski @yanivm13]

@levie Bumped into this tweet - looks cool. curious how you handle state when you go beyond read-only. if Claude iterates on the deck in parallel, do you isolate runs, version, and merge changes? we wrote up a workspace approach for MCP tools here: https://t.co/GSqp31UxVT

[Yaniv Markovski @yanivm13]

@John_Capobianco VibeOps is the right instinct. MCP is a great standard, but reliability lives in “where does this run” and “what state can it touch”. per-task workspaces made that predictable for us. https://t.co/GSqp31UxVT

[Yaniv Markovski @yanivm13]

@DBVolkov skimmed the toc, this looks like a legit curriculum. one add: MCP standardizes how you call tools, but not where they run or how writes stay safe when you go parallel. per-task workspaces helped us a lot: https://t.co/GSqp31UxVT

[Yaniv Markovski @yanivm13]

@neo4j This is great. Curious how you handle isolation when multiple agent runs hit the same Aura Graph Analytics project. do you snapshot, clone, or rely on transactions? we’ve been working on MCP workspaces for this exact “write path” problem: https://t.co/GSqp31U06l

[Yaniv Markovski @yanivm13]

@_akhaliq this is exactly where MCP shines. i made a similar chat-with-arXiv tool with Maestro. once you add “save to doc”, “run experiments”, “edit repo”, isolation becomes the difference between magic and chaos: https://t.co/GSqp31UxVT

[Yaniv Markovski @yanivm13]

@andy_pavlo Loved the “MCP for every database” section. the proxy part is real, but the hard part is safe writes and rollback. we wrote up a workspace isolation approach for MCP style tools here: https://t.co/GSqp31U06l

[Yaniv Markovski @yanivm13]

@VittoStack Anthropic’s Claude Code course is solid. one tip from their docs that’s worth adopting fast: git worktrees for parallel sessions, so agents don’t step on each other. we wrote up the broader “workspaces” pattern here:
https://t.co/GSqp31UxVT

[Yaniv Markovski @yanivm13]

@sahnlam MCP is the standardized tool-calling lane. APIs are the endpoints. for agents that write, the missing piece is “where does this run” + “what state can it touch”. per-task workspaces make it boring but safe.
https://t.co/XNd2TX2I9H

[Yaniv Markovski @yanivm13]

@s_scardapane @iwiwi Love this. AB-MCTS makes TTC an explore/exploit problem: go wider (new candidates) or deeper (refine) from feedback. That's the missing layer for long-horizon agents too: branch, validate, stop early, pick winner. https://t.co/9sWsDgUOzB

[Yaniv Markovski @yanivm13]

@awnihannun Nice demo - budget forcing in the wild. Love the 'Wait'↔</think> knob to dial compute. For tool-using, long-horizon agents, the next layer is orchestrated TTC: branch when needed, validate, stop early, pick best. Deep dive: https://t.co/FdgexsnLJY

[Yaniv Markovski @yanivm13]

@omarsar0 Practical trick is spending compute where it buys certainty: parallel attempts, verification, early stopping. For long-horizon agents that’s orchestration. We shared how this plays out on SWE-bench here: https://t.co/9sWsDgUgK3

[Yaniv Markovski @yanivm13]

Totally. “Tools” is really shorthand for the whole agent runtime. MCP standardizes how you call tools, but it’s silent on where they run, which is fine for read-only tasks and breaks the moment you mutate state (files, docs, downloads, builds). We wrote up a workspace layer for MCP with primitives like initialize/clone/compare/merge/delete so parallel runs don’t collide and you can rollback safely: https://t.co/tuHF2u1SPB

[Yaniv Markovski @yanivm13]

@emollick Our team just posted a series of technical blogs about this topic: https://t.co/FdgexsnLJY

[Yaniv Markovski @yanivm13]

@emollick One year later it’s even clearer, TTC is orchestration, not just “think longer”. rStar-Math shows search+select. For long-horizon agents you want structured TTC: try a few approaches in parallel, auto-kill the dead ends, and stop as soon as one passes the checks

[Yaniv Markovski @yanivm13]

@MillionInt Excellent content

[Yaniv Markovski @yanivm13]

@Mapbox @PointrTech Go Moritz!

[Yaniv Markovski @yanivm13]

We just launched Jamba 1.6 - it outperforms Cohere, Mistral and Llama on key benchmarks, including Arena Hard, and narrows the gap with leading closed models.

Now available on AI21’s Studio and @huggingface
https://t.co/letI7wDL8n