How GC AI Thinks: The Agentic Loops and Graphs Behind Every Answer
Everyone in AI spent months talking about loops. Andrej Karpathy's work at Anthropic popularized the idea that the intelligence in an AI system lives not in the model itself but in the loops built around it. Boris Cherny, who created Claude Code, put it bluntly: "I don't prompt Claude anymore. I have loops running that prompt Claude. My job is to write loops."
Now the conversation has moved one level up. The industry is calling it graph engineering: designing agentic systems as connected workflows rather than one opaque cycle. The loop is not gone. Inside each node, a model still runs the same think-act-observe cycle. Graph engineering is the craft of what happens between those loops: fan-out, fan-in, handoffs, verifiers, and human gates.
GC AI has been building this way since day one. Not for code, but for legal work.
When you upload a vendor MSA into GC AI and type "review this against our playbook," the answer that comes back a few minutes later reads like a senior lawyer sat down and read the whole contract. What you do not see is the work between your question and that answer. Depending on the task, GC AI may have read the document, searched the web for a regulatory update, applied your playbook checks in parallel, verified every citation against the source text, and drafted replacement language for the clauses that missed your positions. All before you finished your coffee.
That chain is not one long chat with a chatbot but a graph of loops in action. An orchestrator coordinates nested agents, parallel checks that fan out and come back as one summary, research branches that fill gaps, and gates where the system pauses for you. Understanding that architecture explains why the output is different from what you get when you ask a general-purpose chatbot to review the same contract.
Loops and Graphs: What They Are
In most AI products, you type a question and get a single response. An agentic loop changes that model. Instead of one pass, GC AI enters a cycle:
Generate a response
If the task requires action (searching, reading a document, running a query), execute that action
Feed the result back to GC AI
Decide what to do next: respond to the user, or take another action
Repeat until GC AI has enough information to deliver a final answer
This cycle is what makes an AI product "agentic." GC AI autonomously decides to take multiple steps to accomplish a task, rather than giving a one-shot answer. And the number of steps it can take, its "loop budget," directly determines the depth and accuracy of the work it can do.
One loop is rarely enough for real legal work. A contract review is not one checklist item but many independent checks, optional research, a synthesis step, and decisions that belong to the attorney. Stuffing all of that into a single context window forces the model to be scheduler, researcher, reviewer, and project manager at once.
So GC AI runs those loops as nodes in a larger graph, each tuned for a different type of legal work. A loop is one agent working toward one goal until it has enough to stop. A graph is several of those loops wired together:
· Parallel edges for independent checks
· Handoffs that return a finding, not the whole scratchpad
· Verification nodes that can catch claims before delivery
· Human gates where your response is the event required to continue
The industry debate sometimes sounds like "loops vs graphs." That is the wrong frame. A loop is a directed cyclic graph with one active node. Graph engineering is what you do when one node is no longer enough. Most of the graph is invisible. From your perspective, you ask a question and get an answer. Behind the scenes, GC AI may have executed 5, 10, or 20+ actions across parallel branches to get there.
Some of those nodes are designed to pause and include a "human-in-the-loop." These human gates are where agentic capability meets attorney oversight. When a task needs multiple inputs to get right, GC AI asks before proceeding rather than guessing. When a review surfaces issues that need your judgment rather than the AI's, that edge of the graph pauses for your input.
The underlying mechanism is the same agentic loop. The difference is that your response is the event required to continue. The AI ran the analysis, but you own the decision.
The Orchestrator Node: Up to 50 Steps Per Turn
Every GC AI conversation starts at an orchestrator node: a main chat loop with a budget of up to 50 loop steps per turn. That orchestrator is the top-level agent that decides which loops to use. Each step can include one or more tool calls (reading a document, searching the web, pulling context from your knowledge base), so the actual number of actions within those 50 steps can be higher still.
Fifty steps is a generous budget, and most tasks do not use all of them. A quick clause explanation might take two steps. A full MSA review using playbooks might chain through fifteen or twenty loop steps.
Those 50 steps are the loop orchestrator's budget for one turn, not the full depth of the work. When GC AI delegates to a research agent, playbook review, or playbook builder, that delegation is typically a single orchestrator step that opens a nested loop with its own separate budget (about 5 steps for research, up to 6 per playbook check in chat, 15 to 30 for playbook building). Parallel agents each run their own nested loop at the same time. What looks like one answer from you can be a handful of orchestrator steps plus hundreds of nested steps inside loops you never have to manage.
This is one of GC AI's real architectural strengths. The orchestrator is the graph’s control node. Nested loops go deep inside each branch. Parallel branches run at the same time. The budgets stack, so the platform never has to trade depth for breadth.
Research Branches: Search, Evaluate, Search Again
When a task requires legal research, web search, or case law analysis, GC AI delegates to a research agent with its own loop. The research agent gets approximately five steps to search, evaluate results, identify gaps in what it found, and search again with a more targeted query.
This loop handles source gathering and refinement. The agent searches, reads the results, decides whether the sources are authoritative enough, and if not, runs another search targeting the gap. A single research question can trigger multiple research agents in parallel, each running its own loop simultaneously across different angles of the same topic.
For in-house lawyers, this means the regulatory summary GC AI delivers has already been through a find-and-refine cycle inside the research agent before it reaches the main chat loop. The research loop's job is to surface the right sources. Whether those sources are accurately reflected in the final answer is a separate verification step, covered below.
Playbook Graphs: Parallel Checks, One Coordinated Run
Playbooks are where the graph architecture delivers the most compound value for in-house teams.
When you run a playbook against a contract, GC AI does not work through your checks one at a time in a single prompt. It fans the review out in parallel. Each check gets its own evaluation loop against the document and the standard or fallback positions you encoded in the playbook. Each check is analyzed against your encoded positions, and GC AI can draft replacement language for clauses that miss. Within each loop, a nested agent can take up to six steps to find the relevant language, compare it to the position, classify the result as pass, fallback, or flag, and pre-draft replacement language when a change is warranted. A 28-check review means twenty-eight of those loops running at once, each focused on one provision, while the overall run stays a single coordinated playbook execution. Suggested fixes come back inside each check result for you to review and apply.
Rachel Harris, General Counsel at Suzy:
"My favorite moment in my career is the day I was able to apply redlines and generate commentary to the opposing party in real time in Word."
The redlines GC AI generated went through that graph before landing in her document. The ones she sent to the counterparty went through her.
A senior lawyer encoded the team's positions once. The graph enforces them consistently, every contract, every reviewer, every time.
Verification Nodes: Accuracy Before Delivery
The layer that separates GC AI from generic AI tools runs after the evidence is gathered but before you see the final answer.
GC AI is designed to verify claims before delivery on substantive work product. When the orchestrator is about to deliver an answer that contains legal authorities, document-grounded claims, calculations, or multi-source synthesis, a verification node can activate. It inventories every high-value claim in the response: citations, quotes, numbers, dates, document references. It checks each one against the tool results GC AI already retrieved, and if a claim cannot be traced back to its source, it makes additional tool calls to verify or correct it. Errors are fixed before the answer reaches you, not flagged for you to catch yourself.
This is the legal-domain version of what Karpathy calls "the gap between 'looks plausible' and 'actually correct.'" In coding, a node can verify by running a test suite. In legal work, the verification gate checks citations against source documents, confirms quoted language character by character, and cross-references claims across multiple sources. The discipline is the same. The stakes, for a lawyer whose name is on the work, are higher.
This verification gate is distinct from the research loop. Research agents gather and refine sources. The verification gate checks whether the final answer accurately reflects those sources. They run at different stages and serve different purposes.
The gate is selective. For casual questions, grammar edits, brainstorming, and short answers with nothing checkable, it skips entirely. It fires when the stakes of getting a claim wrong are real.
On substantive answers, the verification gate is what makes citations trustworthy. Exact Quote lets you confirm in one click: when GC AI says Section 8.2 caps liability at fees paid in the prior twelve months, you click through to the source text and see every character.
Easy Prompt: A Small Loop Before the Graph Starts
Easy Prompt runs a smaller loop of up to three steps when files are attached. Step one: GC AI reads the attached document to understand what type of agreement it is and what it contains. Step two: it maps your plain-language request ("review this MSA for red flags") against the full risk surface for that agreement type. Step three: it rewrites your request into a structured legal prompt that targets the full risk surface (liability, indemnification, IP, data protection, termination, governing law), ready for you to send.
Maury Bricks, General Counsel at ARKO Corp:
"I love how I type in like 'please redline this document' and then press Easy Prompt and it's like, did you mean you wanted to know these 40 things? And I'm like, yes, that's exactly what I wanted to do."
Three steps. Your request becomes the 40-point analysis a senior lawyer would run. You review it and hit send.
What a Full Turn Looks Like
Say your sales team sends you a 30-page vendor MSA at 4 PM and needs the redline by morning. You drop it into GC AI in Word and type "review this against our playbook."
Here is the graph that runs:
· Orchestrator node routes the task
· Twenty-eight playbook-check loops fan out in parallel, each comparing a provision to your encoded position and classifying pass, fallback, or flag
· A research branch can open for a regulatory update referenced in the DPA section
· Results fan back in as one summary
· A verification node can confirm citations before delivery
The map: your message → orchestrator → parallel playbook checks (+ research branch) → verification → human gate (when enabled) → your answer
Loops are how each unit of legal work gets done. Graphs are how those units stay coordinated, parallel, and reviewable. The loops ran. You did the review. The redline goes back that evening.
FAQ
What Is Graph Engineering in Agentic AI?
Graph engineering is the practice of designing agentic systems as connected workflows by wiring multiple loops together as nodes, with defined edges for routing, handoffs, and verification steps. It is the architectural layer above a single loop, organizing how distinct units of work stay coordinated and parallel.
How Is a Loop Different from a Graph?
A loop is a single agent running a think-act-observe cycle toward one goal. A graph connects several of those loops with parallel edges, handoffs between nodes, and human gates so that complex, multi-step tasks can be handled without forcing one agent to act as scheduler, researcher, and reviewer simultaneously.
What Is the Orchestrator Node in GC AI?
The orchestrator is the top-level control node that starts every GC AI conversation and carries a budget of up to 50 loop steps per turn. It decides which nested loops to delegate to, such as research agents or playbook checks, each of which runs its own separate step budget in parallel.
How Does GC AI Handle Parallel Playbook Reviews?
When you run a playbook against a contract, GC AI fans the review out so each check runs in its own evaluation loop simultaneously rather than sequentially. A 28-check review means 28 loops running at once, each analyzing one provision against your encoded positions and pre-drafting replacement language where a clause misses your standard.
What Does the Verification Node Do?
The verification node activates after evidence is gathered but before you see the final answer, checking every citation, quoted passage, number, and document reference against the sources already retrieved. If a claim cannot be traced back to its source, the node makes additional tool calls to correct it before the answer reaches you.
What Is a Human Gate in GC AI?
A human gate is a pause point in the graph where your input is the event required for the workflow to continue. It is how agentic capability is paired with attorney oversight, ensuring that decisions requiring judgment are not made by the AI when they belong to the lawyer.
How Does the Research Agent Work?
GC AI delegates legal research to a dedicated agent with a budget of approximately five steps to search, evaluate sources, identify gaps, and refine the query. Multiple research agents can run in parallel across different angles of the same topic, and their findings are passed to the verification node before reaching the final answer.
What Is Easy Prompt, and How Does It Use Loops?
Easy Prompt runs a smaller loop of up to three steps when files are attached: it reads the document, maps your plain-language request against the full risk surface for that agreement type, and rewrites your request into a structured legal prompt. The result is a comprehensive prompt covering areas like liability, IP, data protection, and termination, ready for you to review and send.
Do I Need to Manage the Graph When Using GC AI?
No. The graph runs behind the scenes, and from your perspective, you ask a question and receive an answer. Depending on the task, GC AI may have executed 5, 10, or 20 or more actions across parallel branches before that answer arrives.
