Glean

Meet Waldo: Glean’s first agentic search model. Built on @nvidia Nemotron 3 Nano and post-trained for search planning, Waldo figures out how to break down a query, which tools to call, what to read next, and when it has enough evidence to hand off.

@glean @nvidia A retrieval planner that runs before the frontier model gets way more interesting at enterprise scale.

One of the questions I often get from leaders is: Can agents reliably get real work done, end to end? A lot of that answer depends on the harness around the model—the system that decides how to break a request into steps, which tools to call, what to remember, and when to stop. Models have a fixed attention span, and the harness decides how it gets filled. Agents are now taking on longer-running, more complex work. To do that reliably and to completion, the harness itself has to be built to scale context. We’ve been solving this problem at @Glean since day one. We wrote up some of what we've learned from our 3rd iteration of our harness here:

@shao__meng @glean Waldo's architecture is clever - separating retrieval planning from reasoning means you don't waste expensive frontier model tokens on mechanical search orchestration. The 10x latency improvement is significant.

Small Language Models are the Future of Agentic AI! Glean just released Waldo - a 30B agentic search model that runs before frontier LLMs. Search is where most agentic work begins. You ask about a project, customer, process, or decision. The agent searches internal docs, reads results, refines queries, searches again. Sometimes one search. Often several iterative loops. Get the search wrong - miss a critical document, surface irrelevant results - and the entire response fails. Search planning is the foundation for AI Agents. But frontier models are doing two different jobs at once. Search planning (which queries, when to stop, is there enough evidence) and synthesis (reasoning over results to generate an answer). The first job is pattern matching. The second needs deep reasoning. Waldo splits these. It's a 30B MoE model built on Nvidia Nemotron 3 Nano that handles just the search planning layer. It runs first, decides which queries to run across Glean Search, Employee Search, and Web Search, determines when it has enough context, then hands off to the frontier model with retrieved context already in place. Key architecture: • Run Waldo first, before the frontier model. The alternative (sub-agent design) would require the frontier model to call Waldo as a tool, wait for results, then respond - two frontier calls. Running Waldo first cuts it to one. • Training Phase 1 (DPO): The model learned when to search, when to stop, and when to hand off from production tool-use patterns. The training data captured which tools were called, in what sequence, and whether the plan succeeded. • Training Phase 2 (RL): The model was trained against production queries and rewarded based on document recall - whether its searches surfaced the same documents that appeared in successful final answers. This refined its ability to find relevant documents in fewer search iterations. • Results: 10x faster per call (250ms vs 3s). Half of queries run on this fast path. The pattern: specialized small models for focused, repetitive tasks. Frontier models for reasoning and synthesis. Waldo proves small language models are faster, cheaper, and just as effective for repetitive, focused tasks. I've shared the link in the replies!

We are in a period of rapid harness iteration driven by a shift in what agents are expected to do. But that's not all. Harnesses are being rebuilt because agents are taking on more work, and more work means more context. That has made us rethink how we design for scale. No longer leaving instructions in the system prompt. Our latest harness redesign reduced our system prompt by 45%+ and focused on four changes: → Sub-agents with isolated context windows → Programmatic tool calling in sandboxes → Compaction strategies → Search-first tool and skill discovery Harnesses are becoming context managers, and getting the design right determines whether an agent can reliably take on work across the enterprise. Read more 👇

Linear Agent x @glean Pull in company knowledge and context from Glean to help Linear Agent generate sharper project plans and updates.

@linear The best work happens when context is already in the flow of work 💯 Excited to bring Glean into @linear Agent!

Lots of great observations in here that match some of our findings with Waldo: glean.com/blog/waldo-lau…

Throwaway line at the end of the convo isn’t being talked about enough! Something to the effect of “We use glean over Claude for my company wide data agent because we trust them with our data” Enterprises won’t 1) single tenant on a lab or 2) give a single lab access to all their data. Still lots of spend up for grabs to be the all knowing, all access AI like Ev described. Also I wouldn’t be too worried about Claude replacing O365/g suite anytime soon!

@nvidia We think enterprise AI will need specialized models alongside specialized tools — especially for high-volume jobs where latency and cost matter. Thank you to @NVIDIA and @thinkymachines for their partnership on Waldo!  Read more: glean-it.com/4efyJX7

Not every enterprise query should hit a frontier model first. For retrieval-heavy work, that’s often the slowest and most expensive path.  Waldo runs first, builds context, and helps route the job to the right reasoning level. From evals: 10x faster on a per-LLM call basis, translating to ~50% lower latency and ~25% lower token usage when integrated into our harness, with no quality regression.

Meet Waldo: Glean’s first agentic search model. Built on @nvidia Nemotron 3 Nano and post-trained for search planning, Waldo figures out how to break down a query, which tools to call, what to read next, and when it has enough evidence to hand off.

@thewyattbrocato @pmitu @perplexity_ai

@pmitu Currently Claude is my favorite to use… but not on their platform. Why run out of credits when I can use a tool like @perplexity_ai or @glean and use Opus or Sonnet freely

If I had to choose 1 AI model, I'd pick Claude.

GPT-5.5 from @OpenAI is here! In early evals on enterprise workflows, we’re seeing promising results on instruction following, latency, and token efficiency. Coming soon to Glean Assistant and Agents 👀

@silvercorp Come say hi 👋, we’ve got some great demos lined up!

Today I'm going to check out the @glean section. I'll let you know what demos they have. It's one of my favorite and most useful tools for day today work 😀 #GoogleNext

@RebHinds Watch the episode: glean-it.com/3OHhUKd

How do you actually move AI adoption forward? In this episode of Intelligence: Real & Imagined, @RebHinds, Shweta Puri, and Nichole Sterling get into what actually makes AI stick: start with one workflow, make it smarter, and build from there.

@JayaGup10 @jainarvind glean-it.com/3ONnVFk

The reality in enterprise AI = context > models @JayaGup10 & @jainarvind dig into this topic and where AI breaks down in practice, who should own the context graph, and whether context graphs are really a trillion-dollar opportunity. Watch the full conversation linked below 👇