AI Engineer World's Fair 2025
3 ingredients for building reliable enterprise agents
Loading slides…
Read the talk
Three ingredients for reliable enterprise agents
Enterprise agents become useful when they do valuable work predictably, make mistakes recoverable, and keep people involved at the decisions that matter.
From a talk by Harrison Chase
Before you start: Familiarity with LLM tool calls, retrieval-augmented generation (RAG), and Git branches and pull requests will help with the examples.
An agent for every task—and a human supervising
What would it take to run an enterprise full of task-specific agents, with people coordinating their work as managers and supervisors? The challenge applies both to developers building agents inside a company and to vendors selling agents into one. A compelling vision of delegated work still has to become something an organization will adopt.
Why do some agents succeed in that setting while others fail? Harrison Chase credits a conversation with Assaf, whom he introduces as Monday’s head of AI and the author of GPT Researcher, for many of the ideas behind his answer. The starting point is the economics of using an agent, before choosing its architecture.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The production threshold
An agent becomes more attractive as the value when it succeeds increases. It also becomes more attractive as its probability of success increases. The third ingredient moves in the opposite direction: a high cost when it fails discourages adoption. These are separate levers. A valuable task can justify imperfect performance when mistakes are cheap; a costly mistake can make an otherwise capable agent difficult to approve.
Chase expresses the relationship as an expected-value threshold. Let p be the probability of success, V the value of a successful result, L the loss from failure, and C the cost of running the agent:
The same reasoning applies beyond enterprises. The practical question is how product design and engineering can move these terms enough to make deployment worthwhile.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Increase the value of the work
The most direct lever is task selection. Chase points to Harvey in legal work and to research, investment research, and summarization in finance. People already pay substantial amounts for lawyers and professional research. That existing willingness to pay makes the value of a successful result easier to identify.
But increasing value does not require changing industries. It can mean changing how much work the agent completes per interaction. Chase describes older RAG and question-answering interfaces as aiming to answer within five seconds. Deep research instead spends an extended period investigating a question. In coding, Cursor’s inline autocomplete and chat illustrate short interactions; background coding agents represent a larger unit of delegated work. Chase informally counts seven examples appearing in the preceding three weeks, with agents running for hours at a time.
The coding slide makes the contrast concrete with the labels “Cursor: 30 seconds” and “Claude Code: 2 hours.” These illustrate different work horizons, rather than a controlled speed comparison. The product opportunity is to let an agent do more substantial work autonomously in the background. That requires changing the UI and interaction pattern, not merely replacing the model behind a quick-response chat box.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Put required behavior in control flow
Increasing the probability of success starts with the gap between a prototype and a production system. A run that works once can make a good demonstration without establishing reliability. Some applications tolerate exploratory behavior whose steps are difficult to predict. Enterprises often need tighter control: a particular step must follow another, regardless of what the model would otherwise choose.
In Chase’s illustrative example, prompting an agent to perform A after B might preserve that order about 90% of the time. Encoding the sequence in deterministic control flow makes the ordering a property of the program. It does not guarantee that either step produces a correct result; it removes one decision from the model’s discretion. More of the agent can become deterministic wherever predictability is a requirement.
Anthropic’s Building effective agents distinguishes workflows from agents. Chase’s preferred framing is workflows and agents: tool-calling loops can coexist with fixed sequences. If agent A always hands its result to agent B, the outer structure is a workflow even when each agent makes its own internal decisions. In TypeScript, that boundary can be as small as this:
typescript
type Agent<Input, Output> = (input: Input) => Promise<Output>;
async function runInOrder<Input, Intermediate, Output>(
input: Input,
agentA: Agent<Input, Intermediate>,
agentB: Agent<Intermediate, Output>,
): Promise<Output> {
const resultA = await agentA(input);
return agentB(resultA);
}
Here, agentB receives the completed result of agentA; if A throws, B is not invoked. The agents can remain flexible internally while their handoff stays explicit.
LangGraph is Chase’s example of a framework designed around this spectrum. The useful question is where an application needs deterministic structure and where it benefits from model-directed execution. Different applications need different positions on that spectrum.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make performance legible to reviewers
Actual reliability is only part of getting an agent approved. Stakeholders may also have substantial uncertainty about how it behaves. Improving the agent and reducing that uncertainty are related but distinct jobs. Observability and evaluations help with both: they support engineering work while making performance understandable to the people deciding whether deployment is acceptable.
Chase describes LangSmith as a tool originally built so developers could inspect their agents, which also became useful for communicating with external stakeholders. Step-level traces reveal where a run succeeds, where it fails, and which patterns recur. A reviewer can see several LLM calls and the steps connecting them instead of imagining an opaque, single-shot answer. Evaluations and benchmarks add evidence about how the system performs.
Chase recounts a customer who first used LangSmith to build an agent, then brought the same view into a production review meeting. Showing the panel what happened inside the system helped reduce perceived risk, and the meeting finished early. The anecdote illustrates a second audience for developer tooling: the people responsible for approving its use.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Reduce the cost of being wrong
Reviewers also worry about consequences: an agent could damage a brand or give something away for free. Chase argues that those fears can dominate the adoption decision. Two product mechanisms address different parts of the risk: make changes reversible, and place a human approval boundary before consequential actions.
Code offers a natural example of reversibility. Chase points to a Replit Agent screenshot while discussing generated diffs and pull requests, then explains how returning to a previous commit can undo code changes. In his historical account, Replit saved each file change as a new commit. Today, Replit’s checkpoint documentation describes checkpoints at milestones and other strategic moments, rather than every edit. The enduring mechanism is a recoverable earlier state; it does not imply that every external effect can be rolled back.
Approval and reversal are separate protections. Opening a pull request instead of merging directly into main lets a person review the proposed change before it reaches the main branch. Reversal provides a recovery path after a change; approval provides a chance to stop it beforehand. Together they change the consequences of an agent mistake—and how an enterprise evaluates those consequences.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Clarify intent, then produce something reviewable
Deep research shows how human involvement can increase value as well as limit risk. Before investigating, the system asks follow-up questions and calibrates what the user wants. That exchange improves the chance that the eventual report addresses the right problem. At the other end, the output is a report for the user to read and act on—not an automatically published blog post or an email sent to clients. Research is delegated; publication remains a separate decision.
Chase describes a similar pattern with Claude Code: clarifying questions help establish intent, and a separate branch plus a pull request preserves a review boundary. The workflow does not require a commit for every individual file change. It requires keeping proposed work separate from the main branch until someone decides to accept it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Scale useful work beyond the chat box
Once an agent has positive expected value, the next opportunity is to repeat that useful work at greater scale. Ambient agents run in the background and start in response to events, rather than requiring a person to initiate every run. Chase contrasts managing one or two chat boxes with potentially hundreds of agents working in the background. The interaction changes from one-to-one to one-to-many. His description of unlimited concurrency concerns that interaction model, not measured infrastructure capacity.
The latency expectation changes too. In chat, someone is waiting for a response. With event-triggered background work, there is less pressure to return an immediate answer. That creates room for longer sequences of reasoning and tool use: changing a line of code can become changing a whole file or creating a repository. More time is useful because it permits a larger body of work, rather than merely a slower answer.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Background execution still needs oversight
Ambient does not mean fully autonomous. Running in the background does not remove the need for an interface through which people can supervise the work. Otherwise, the same fear returns: the agent might act without anyone understanding or controlling the consequences.
Chase describes several intervention patterns:
- Approve or reject: Require explicit permission before selected tools can run.
- Edit a tool call: Correct the proposed action in the UI instead of discarding the entire run.
- Answer questions: Let the agent ask for clarification or missing information when it gets stuck.
- Revisit an earlier step: After a run, return to a prior point and resume with different guidance. Chase calls this time travel, or human on the loop; his example is returning to step ten of a hundred after identifying a mistake there.
These controls place human judgment at particular boundaries without requiring a person to direct every internal operation.
There is also an intermediate stage between chat and ambient agents. Under Chase’s stricter definition, deep research and Claude Code are not ambient because humans still initiate them. He likes Factory’s term, async coding agents, for work that starts with a human request and continues in the background. Initial interaction calibrates the task; execution then proceeds asynchronously, with opportunities to ask for help.
Chase therefore adds a conceptual middle column to the chat-versus-ambient distinction:
| Interaction | Starts with | Work pattern |
|---|---|---|
| Chat | Human message | Immediate exchange |
| Sync-to-async | Human request and calibration | Background execution |
| Ambient | External event | Background execution with selective oversight |
The key distinction is who or what starts the work. Asynchronous execution alone does not make an agent ambient.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An inbox for pending agent actions
Agent Inbox gives this supervision model a concrete interface. It surfaces actions that need human attention so the user can approve, reject, or leave feedback. Those are interaction descriptions: the inspected implementation names its responses accept, edit, response, and ignore. A pending action is something the agent wants to do, not evidence that it has already done it.
Email is a natural application. An incoming message is an event that can initiate work without a user opening a chat. Many incoming messages can create concurrent tasks, while outgoing emails or calendar invitations remain subject to approval according to the user’s comfort level. Event-driven intake and controlled external actions can therefore coexist.
Chase built and uses an email agent to test these patterns internally. He closes the presentation by offering its open-source implementation through an on-screen QR code. The example brings the architecture together: events start the work, the agent prepares actions, and an inbox provides a place for human decisions.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What generalizes beyond coding agents
In the Q&A, an audience member asks whether coding agents attract funding because their work is measurable and reversible. Chase answers by separating two advantages of code, without establishing the question’s premise that only coding agents receive funding.
The first advantage is verifiability. Code can be run and checked for compilation, although compilation alone does not establish correctness. Math also offers results that can be checked. Essay quality is more ambiguous: deciding whether an essay is correct is a less sharply defined task. Chase connects this difference to training—verifiable outputs make it easier to bootstrap training data, which contributes to stronger coding models and, in turn, stronger coding agents.
The second advantage is a product pattern that travels more easily across domains: first drafts. Code naturally supports commits, drafts, and previews, but legal documents and essays also have familiar draft-and-review workflows. An agent can complete a substantial amount of work before presenting an artifact for human judgment.
Putting a human in the loop at every tiny step erodes the value of delegation. The design challenge is to let the agent do substantial work while keeping people involved at key points. A first draft supplies exactly that boundary: enough completed work to be useful, with a clear decision before acceptance or use. That pattern generalizes across legal work, writing, and code more readily than the verifiability that helps train coding models.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Anthropic's guide to workflows, agents, and combining simple architectural patterns.
Framework for stateful workflows and agents with persistence and human oversight.
An inbox interface for reviewing agent actions, editing arguments, and supplying feedback.
An open-source agent for conducting research using language models.
Further reading
- Introducing ambient agentsArticle
Harrison Chase explains event-triggered agents, selective human oversight, and an email-assistant example.
Email-assistant implementation with triage, drafting, calendar configuration, and Agent Inbox integration.
Updates since the talk
Current checkpoint creation and rollback behavior, including database restoration limits.
Full transcript
Loading transcript…