How Atlas Agents execute document review and knowledge graph curation at scale

Atlas Agents: the execution layer of AtlasAI Singularity

AtlasAI Singularity (v6) ingests the firm's documents from iManage, SharePoint, OneDrive, NetDocuments, and email into a per-tenant knowledge graph. But ingestion is just the starting point. The graph is curated continuously by Atlas Agents, which read, classify, link, and refine new content as it lands. This continuous curation is what makes every Atlas response grounded in the firm's own work, not in generic LLM knowledge.

Atlas Agents ship as the execution engine for that curation, and also as the work engine for document review workflows inside Workspaces. They handle multi-step orchestration, parallel document processing, and the generation of new client documents from review findings.

Two jobs: Workspace execution and knowledge graph curation

Inside a Workspace (a matter-scoped environment), agents execute document review workflows at scale. When a team uploads hundreds of documents, agents run in parallel to:

• Classify and route documents (by type, counterparty, risk level)
• Execute pre-review processes (privilege checks, conflict scanning, scope validation)
• Run substantive review (extract clauses, flag R&Ws, spot indemnification language, surface diligence findings)
• Draft new client documents from the review output (SPA markups, indemnification schedules, diligence memos, closing checklists)

These aren't prompted workflows each time. They're designed once, validated, and reused on every new matter. An M&A team builds a Deal Review agent; an IP team builds a Patent Diligence agent; a Litigation team builds a Privilege Review agent.

Outside Workspaces, the same agent execution layer curates the knowledge graph itself. As new documents land in firm systems, agents ingest and refine them. They classify content, extract key facts and relationships, link documents to matters and precedent, and keep the graph current. This continuous curation is silent but critical: it's what ensures that when a partner searches for "comparable M&A indemnity caps" or "IP diligence templates," they're querying a sharp, firm-specific graph, not a generic corpus.

What shipped: planner execution and PDF handling

This week we shipped critical stability fixes for agent execution at scale. Agents use a planner component to decompose complex workflows into multi-step tasks. We increased the token budget for planner JSON parsing and added truncation recovery, so agents can handle larger review datasets without running out of space mid-workflow. We also shipped defensive JSON serialization for agent task boards, fixing serialization errors that would crash the board view when tasks or results contained edge-case data.

We also fixed PDF rendering in the workspace viewer to handle any file type cleanly, and added inline review editor support so agents can reference, edit, and update findings without leaving the interface.

Concrete workflows: how teams use agents

M&A Deal Review: Upload 200 purchase agreement precedents and the target company's cap table. Agents classify agreements by counterparty and deal stage, extract material terms (R&Ws, indemnification, escrow, reps, survival), compare them to the firm's historical playbooks in the knowledge graph, and draft a marked-up SPA for the new transaction. Output is tabular reviews showing each extracted clause and a marked summary document ready for partner review.

IP Patent Diligence: Upload a portfolio of competitor patents and the firm's own prior art. Agents extract claims, classify by technology area, compare claim language to firm precedent, flag potential invalidity and infringement risks, and produce a diligence memo with rankings. All structured in a tabular review so the team can drill into claim language, prior art references, and agent findings side by side.

Litigation Privilege Review: Upload a corpus of emails and documents from a litigation hold. Agents scan for privilege markers (attorney work product, legal advice, confidentiality), route flagged documents to a privilege review folder, and draft a privilege log with attorney names, dates, subjects, and privilege assertions. The knowledge graph learns which sender pairs, subject lines, and terminology patterns indicate privileged content, so privilege detection improves with each review.

Why agents matter for knowledge graph grounding

Without agent curation, a firm's knowledge graph is a static snapshot. With agents, it's alive. New contracts land in iManage; agents read them, extract key terms, and link them to related matters and precedent. Attorneys update playbooks; agents surface the changes to team members working on similar deals. A litigation search for "class action indemnity language" doesn't return raw documents; it returns agent-curated insights tied to specific claims, defenses, and settlement patterns.

This grounding is why agent-driven document review is faster and more accurate than generic prompting. Agents work against a graph that's shaped by the firm's own work, not a generic corpus. And the more the firm uses agents, the sharper the graph becomes.

Getting started with Atlas Agents

Start in a Workspace. Upload documents related to a real matter, then build or run an existing agent action list. Watch agents classify, review, and draft in parallel. Check the agent task board to see what each agent found, edit results inline, and download the drafted client documents. Build your first reusable agent action list from successful one-off runs, then reuse it on the next matter.

Visit https://atlas-ai.io to see agents in action.