Legal ops · Production

Coinbase builds enterprise AI agents: six-week sprint to standardize agentic automation for internal workflows

The problem

Coinbase internal teams ran manual, time-consuming, decision-heavy workflows in Institutional support, Onramp onboarding, and Listing legal review, with no standardized way to build, host, or audit AI agents at enterprise scale.

First attempt

Low-code tools proved unsuitable for long-running operational flows because loading more tools and instructions into a prompt introduced context noise, making outputs harder to reproduce and individual steps harder to unit test or gate in CI.

Workflow diagram · grounded in source
1
Manual work identified in automation domains
trigger
“We focused on building out process automation agents in the Institutional support, Onramp onboarding, and Listing legal review spaces”
2
LLM nodes run within code-first graph
ai_action
“Code‑first graphs (e.g., LangGraph/LangChain patterns) gave us typed interfaces, version control, clean separation of “data” nodes from “LLM” nodes”
3
Evaluation harness and LLM-judge validation
validation
“LLM steps run with evaluation harnesses and curated datasets. We use a second LLM as a judge for spot‑checks and confidence scoring”
4
Human review of agent output
human_review
“we treat human review as an intentional part of the system, not a workaround”
5
Immutable audit record created
output
“An immutable record showing which data was used, how it was used, the reasoning the agent followed, and who approved the output is created for each execution of an agent”
Reported outcome

Two automations reached production with more than 25 hours saved per week; new agent build time fell from 12+ weeks to under a week, and more than half a dozen engineers were able to self-serve on the patterns.

Reported metrics
Automations in productiontwo
Weekly time saved from production automationsmore than 25 hours saved per week
New agent build time reductionfrom 12+ weeks to under a week
Time to build a new agentdropped from quarters to days
Show all 5 reported metrics
automations in productiontwo
weekly time saved from production automationsmore than 25 hours saved per week
new agent build time reductionfrom 12+ weeks to under a week
time to build a new agentdropped from quarters to days
engineers self-serving on patternsmore than half a dozen engineers
Reported stack
LangGraphLangChainLangSmithPython
Source
https://www.coinbase.com/en-nl/blog/building-enterprise-AI-agents-at-Coinbase
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Two automations reached production with more than 25 hours saved per week; new agent build time fell from 12+ weeks to under a week, and more than half a dozen engineers were able to self-serve on the patterns.

What tools did this team use?

LangGraph, LangChain, LangSmith, Python.

What results were reported?

Automations in production: two; Weekly time saved from production automations: more than 25 hours saved per week; New agent build time reduction: from 12+ weeks to under a week; Time to build a new agent: dropped from quarters to days (source-reported, not independently verified).

What failed first in this deployment?

Low-code tools proved unsuitable for long-running operational flows because loading more tools and instructions into a prompt introduced context noise, making outputs harder to reproduce and individual steps harder to…

How is this legal ops AI workflow structured?

Manual work identified in automation domains → LLM nodes run within code-first graph → Evaluation harness and LLM-judge validation → Human review of agent output → Immutable audit record created.

Related legal ops cases