Enterprise agents need to create substantial value when they work, while keeping the cost of mistakes low.
2
Reliability comes from choosing where to use deterministic workflows alongside agentic loops, then measuring and exposing how the system behaves.
3
Ambient agents can run in response to events and handle more work, but they still need human approval, correction, and review points.
Summary
Harrison Chase frames enterprise agent design around three factors: the value produced when an agent succeeds, the probability that it succeeds, and the cost when it fails. High-value tasks include legal work, investment research, deep research, and longer-running coding tasks. Reliability improves when developers make predictable parts deterministic instead of relying on prompts for every step. Chase describes LangGraph as a framework for combining workflows and agents along that spectrum. Observability and evaluation also help review boards understand what an agent does and where it fails. To reduce the cost of mistakes, agents should create reversible changes and involve people at approval points. Chase then describes ambient agents, which respond to events and run in the background. They can handle more work than chat agents, but ambient does not mean fully autonomous. Approval, editing, questions, and time travel remain part of the interaction.
Enterprise agents should target work with high value when they succeed
Chase starts with the value an agent provides when it is right. He points to Harvey in legal work and finance agents for research and summarization because people already pay heavily for those tasks. Value can also rise when an agent does more substantial work over a longer period. Deep research systems run for an extended time instead of returning a quick answer, while newer coding agents can work in the background for hours. Chase says the future he is considering involves agents working more autonomously in the background, rather than acting only as copilots.
Making predictable steps deterministic raises an agent's reliability
A prototype can work once and still be difficult to operate in production. Enterprise users often want to know that step A will always happen after step B. Prompting may produce that sequence about 90% of the time, but code or a deterministic workflow can guarantee it. Chase describes the relationship between workflows and agents as a spectrum. Some parts loop, call tools, and make decisions, while other parts simply execute A, then B, then C. Multi-agent systems can occupy the middle ground when one agent always calls another after it finishes.
Observability reduces uncertainty around an agent's behavior
Chase says enterprise teams face large error bars when they try to approve new agent technology. LangSmith was built so developers could inspect what happens inside an agent, but it also helps explain the system to external stakeholders. People can see every step, including whether the agent makes three or five language-model calls, and can benchmark its performance. Chase describes a user who showed LangSmith traces to a production review panel. The panel finished early because it could inspect what the agent did and where it failed instead of discussing an opaque system.
Reversible changes and human approval lower the cost of mistakes
Agents are easier to approve when their actions can be undone. Chase uses code as an example: Replit Agent creates commits for file changes, so users can return to an earlier state. A human can also review a pull request before code reaches the main branch. These controls change how people calculate the consequences of a bad action. Deep research follows a similar pattern by asking questions up front and producing a report for the user to review, rather than publishing or emailing the result automatically. Claude Code also asks clarifying questions before acting.
Chat agents are limited by the number of conversations a person can manage and by the expectation of a fast response. Ambient agents respond to events and run in the background, so many can operate at once. Their looser latency requirements let them perform larger operations, such as changing an entire file or creating a repository instead of making one tool call. Chase says ambient agents still have a user experience. An agent inbox can collect actions that need approval, and users can approve, reject, or leave feedback.
Chase rejects the idea that a background agent must run without oversight. Possible interaction patterns include approving or rejecting particular tool calls, editing a tool call, answering questions when the agent gets stuck, and using time travel to resume from an earlier step with different instructions. He describes deep research and coding agents as a transition from synchronous to asynchronous work because a person starts them and calibrates the task, while the agent can continue working before asking for help.
First drafts are a general interaction pattern beyond coding
In the question period, Chase explains why coding agents have attracted attention and funding. Code and math are verifiable because systems can compile code or check an answer, which gives model builders training data. Code also supports commits, drafts, and previews that make changes easy to review and reverse. He argues that the first-draft pattern applies to legal documents and essay writing as well. The agent can do substantial work, while the human reviews a draft at a meaningful point instead of approving every tiny step.
"The core thing that we've seen is that it's easy to get something that works in a prototype. It runs once great. You can make a video, put it on Twitter, but it's hard to make it work reliably, put it in production."05:30
Who should watch
You are building an enterprise agent and need a way to reason about adoption, reliability, and failure costs before deployment.
Your agent has predictable steps mixed with open-ended decisions, and you need guidance on where to use workflows and where to keep agentic behavior.
You are designing background agents and need approval, correction, and review patterns that keep people involved without checking every small action.
