Glean's $300M ARR Horizontal Neo-Lab: Context-First AI, Trace Learning & Waldo Model

Disruptive Evolution Sequence: Custom Model Comes Last

A common enterprise AI startup path is: fine-tune an open-source model into an "industry LLM" → wrap an Agent shell for demos → sell to vertical customers. Glean's public product and technical materials reveal a different sequence:

1. Enterprise system connectors & permission alignment (Connectors & Permissions)
↓
2. Enterprise relationship graph & context foundation (Enterprise Graph & Context)
↓
3. Enterprise-grade Agent orchestration skeleton (Harness & Sandbox)
↓
4. High-value traces from real business execution (Production Trace)
↓
5. Private evaluation & experience distillation (Private Eval & Trace Learning)
↓
6. Only then launch post-training for high-frequency stable subtasks, releasing custom model Waldo

The article frames this as an application-type Neo-Lab: enter real enterprise work, accumulate context and execution traces, then train specialized models for closed-loop stable tasks.

Why This Is No Longer RAG: Controlling the Enterprise Context Layer

Traditional RAG works via: user query → vector retrieval → assemble document chunks → LLM summarizes. As base model context windows expand and vendors add search plugins, such solutions become easily replaceable. Glean's scope includes an Enterprise Context Platform :

Enterprise Context
= Deep Connectors
+ Hybrid Search
+ Real-time Permissions
+ Enterprise Relationship Graph (People ↔ Content ↔ Activity ↔ Org)
+ Work Memory & Action Catalog

Document retrieval is only one problem. The system must also handle:

Permissions & compliance (ACL) : interns cannot see financial reports; sales cannot see HR salaries.

Dynamic organizational relationships : who reports to whom, who recently had high-frequency activity on a project, which old documents remain most authoritative.

These permissions and org relationships come from the enterprise itself. Models can be swapped, but the Context foundation must continuously connect, sync, and inherit the enterprise's real access boundaries.

Harness Engineering: Moving Working Memory Out of the Context Window

In complex enterprise tasks, stuffing all information into the LLM context window increases context pollution, attention degradation, and token cost . Glean delegates this orchestration to a Harness (execution framework) :

Sandbox as short-term working memory : gives each Agent an isolated filesystem, CLI, and Python runtime. Massive intermediate data and raw logs stay in the Sandbox; the LLM context window retains only the current core decision state.

Programmatic Tool Calling : abandons inefficient multi-round JSON tool calls; lets the model generate Python code to batch-call enterprise APIs inside the sandbox. Glean officially disclosed the new Harness reduces overall token consumption by ~ 24% (not solely attributable to code-style calling).

Tool Search : large enterprises have thousands of internal APIs and MCP tools. Harness performs tool retrieval at task initiation, so the model sees only a filtered minimal candidate set, lowering mis-invocation rates.

Trace Learning: Turning Experience into Assets

Enterprise task success often depends on implicit business rules : which tools must be called in what order, which fields need cross-system verification, who to escalate conflicts to. Glean addresses this with a Trace Learning loop :

Real business execution produces Production Traces
↓
Offline evaluation: strong model (Teacher) replays multiple candidate paths, compares correctness, efficiency, tool behavior
↓
Experience distillation: extract natural-language "Strategy Memory"
↓
Online injection: next similar task, Harness auto-retrieves and injects strategy memory to guide execution

This is called Harness-level Continual Learning . Glean's public statements indicate these learned strategies are stored outside the model, isolated by deployment and user. This avoids writing every business feedback directly into model weights and reduces catastrophic forgetting risk. Specific enterprise compliance conclusions still depend on data handling, permissions, and deployment modes; they cannot be inferred solely from this mechanism.

Waldo: When Should a Custom Model Appear?

In 2026, Glean launched its custom small model Waldo (fine-tuned from NVIDIA Nemotron 3 Nano, MoE architecture with 30B total parameters, ~3B active parameters).

Waldo does not produce final answers; it handles "Agentic Search Planning" (search planning & task routing). It solves one problem: given a complex question, how to decompose search intent, which tool to call, how many searches, when to stop, when to hand off to a frontier LLM.

Why this task? Because it exhibits three traits:

Ultra-high invocation volume (High Volume)

Extremely well-defined task boundaries (Well-defined)

Highly objectively verifiable results (Verifiable)

Per Glean's public intro, Waldo uses anonymized behavioral traces for DPO (Direct Preference Optimization) and RL, not customer document content. Internal benchmarks show:

Single call 10x+ faster than default reasoning model.

Overall end-to-end latency reduced ~50% .

Token usage cut ~25% with no observed quality drop.

These are Glean's internal test results, not third-party production audits. They illustrate the right place for specialized small models: handle high-frequency, well-bounded, continuously evaluable subtasks, leaving expensive frontier models for final reasoning.

Private Eval & Model Routing: Making Base Models Swappable Capabilities

Glean's Model Hub integrates 40+ open-source and commercial frontier models; platform capabilities are not locked to any single vendor. In August 2026, Glean published an analysis covering ~1,000 enterprise tasks across 37 models and inference configurations , and states it runs thousands of evaluations per quarter.

Based on this Private Eval suite, Glean performs two-layer dynamic routing at runtime:

Model Selection : for this task, which performs best — open-source small model, Claude, or GPT?

Effort Routing : how much Chain-of-Thought (CoT) depth does this task need?

This design lets the system swap base models per task. The enterprise's long-term Context, Harness, Trace, Eval, and Memory constitute the harder-to-migrate system assets.

Two Giants Side by Side: Glean vs. Harvey Roadmap Comparison

Entry Point
Harvey (Vertical Neo-Lab): Legal top-level workflows
Glean (Horizontal Neo-Lab): Enterprise-wide search

Core Domain
Harvey: Legal / professional services compliance
Glean: Enterprise operations & collaboration

Context Foundation
Harvey: Matter files / regulation libraries / Vault
Glean: Relationship graph (Enterprise Graph) / permission model

Implicit Knowledge Source
Harvey: Practicing lawyers & Legal Engineers (human distillation)
Glean: Company-wide collaboration & tool-call traces (system extraction)

Custom Model
Harvey: Tenet (complex legal analysis & generation)
Glean: Waldo (high-frequency search planning & tool dispatch)

Evolution Direction
Harvey: From single Vertical Job toward underlying platform
Glean: From Horizontal platform toward Vertical Jobs

Both routes combine the same system capabilities:

Frontier Models
+ Specialized Small Models
+ Private Eval
+ Enterprise Harness & Memory
+ Enterprise Context Layer

Five Takeaways for Domestic Enterprise AI Founders

If you are landing AI in enterprises (manufacturing, retail, finance, supply chain), Glean's path offers five engineering trade-offs:

Solidify connectors & permission control first : Agents need latest, permission-verified ground-truth data; otherwise answer quality and access boundaries stay unstable.

Retain full execution traces (Trace) from day one : beyond chat logs, record task goals, tool calls, parameters, errors, and final human corrections. Only this supports later evaluation and strategy extraction.

Treat Harness as an independent engineering system : sandbox isolation, code-style invocation, tool search, and state management jointly determine stable execution; prompting alone cannot bridge gaps.

Prioritize retrievable strategies at the Harness layer : enterprise knowledge changes fast and is permission-sensitive. Storing strategies outside the model enables updates, isolation, and auditability; whether to adjust model weights is decided per task and compliance needs.

Start training small models on high-frequency, well-defined, verifiable subtasks : routing, screening, and planning have clear boundaries and easier evaluation standards. Waldo demonstrates how targeted training reduces cost and latency.