AI Engineer World's Fair 2025

Mentoring the Machine

Eric Hou19:06

Loading slides…

Read the talk

Mentoring the Machine: From Interrupted Work to Parallel Exploration

An overdue component, a staging failure, and an onboarding question show how agents change engineering work when people supply context, set boundaries, and evaluate results.

From a talk by Eric Hou

Before you start: Familiarity with software repositories, incident response, and code review will help; no prior experience with Augment is required.

One engineer, three competing demands

How do you finish an overdue design-system component when staging breaks and a new hire needs your help? For Eric Hou, a member of technical staff at Augment Code, that question belongs to the everyday reality of production engineering. Before Augment, he spent six years building automotive products and standards. He describes systems touched by tens of thousands of engineers, where no individual fully understood even five percent of the whole. Work in that environment depends on finding knowledge distributed across people and systems.

His Tuesday begins at nine in the morning with a design-system component that should have merged the previous Friday. Design and downstream teams are waiting. At nine thirty, a client/server request-format mismatch breaks the main staging API endpoint, blocking QA and deployments. The primary owner is on vacation; Hou is the service secondary and must handle the incident. At ten fifteen, while he is investigating service logs, a new hire asks for help understanding the extension system. The immediate demands now compete for the same attention, and incident remediation threatens to add another week or two of work.

Slide with three cards: 9AM Design System Component, already late with the team waiting; 9:30 AM Staging Broken, request format mismatch blocking all testing; and 10:15AM New Hire Engineer, needing guidance for an onboarding project.Open full source frame
“One of those days”: a delayed design component, broken staging, and a new hire needing guidance.

Hou motivates the problem with figures of 23 minutes to recover from an interruption, two-thirds of engineering time spent maintaining code, and $300 billion annually lost to context switching and firefighting. These are his cited figures, without measurement conditions established in the talk; they should not be treated as universal recovery costs or a measured baseline for his experience. The practical problem is recognizable without them: urgent work interrupts planned work, and the follow-up creates another backlog.

0:200:33
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:20 · section reference included

Give the agent an outcome and boundaries

Before replaying how Tuesday actually unfolded, Hou starts a task in the Augment Agent. He asks it to create a personality reflecting the energy of the AI Engineer World Fair and San Francisco. The prompt supplies the desired behavior, followed by guidelines that define the agent’s boundaries. It does not prescribe every implementation step. He starts the agent and returns to the presentation, leaving the feature work running in the background.

Delegation begins with enough structure to work independently. The agent receives an outcome and constraints; the human remains available to evaluate what comes back. Running the task during the presentation makes the overlap visible: implementation can continue while the person who requested it does something else. That same pattern changes the shape of Hou’s interrupted Tuesday.

4:014:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:01 · section reference included

Keep work moving while handling the incident

In the actual morning, Hou scopes the design-system component with an agent before getting coffee. He provides outcomes, context, and constraints, then asks it to do the preparatory work he would expect from another engineer: explore the codebase and draft an RFC. When he returns, the RFC is mostly complete and follows the existing architectural patterns. The agent has used the repository to turn a requested outcome into a proposed implementation direction.

When the staging emergency arrives, Hou keeps the component moving in the background and separates the incident work into parallel activities:

  • Implementation: An agent continues the design-system component.
  • Diagnosis: Two agents help analyze logs and perform git bisect.
  • Communication: The Augment Slack bot helps update teams that cannot deploy.
  • Coordination: Hou concentrates on the critical path of resolving the incident.

The work still needs engineering judgment, but each investigation no longer requires him to stop every other activity.

The git bisect work provides a concrete example of a bounded diagnostic task. Given a known-good revision, a known-bad revision, and a repeatable check for the request-format failure, Git can narrow the regression to a commit. A shell script for that workflow can accept the revisions and test command as arguments:

bash

#!/usr/bin/env bash
set -u

if (( $# < 3 )); then
  echo "Usage: $0 GOOD_REV BAD_REV TEST_COMMAND [ARGS...]" >&2
  exit 2
fi

good_rev=$1
bad_rev=$2
shift 2

git bisect start "$bad_rev" "$good_rev" || exit 1
trap 'git bisect reset' EXIT
git bisect run "$@"

The check must return zero for a good revision, a failure status for a bad one, and 125 when a revision cannot be tested. Defining that check is part of supplying context: the agent needs a precise way to recognize the failure, not merely a request to find something suspicious in the history.

At ten fifteen, the onboarding question takes a different route. Hou directs the new hire to the Slack bot, which has access to Augment’s context engine, codebase, documentation, and Linear. The new hire can ask follow-up questions and receive help grounded in the organization’s own material while Hou stays focused on incident response. This is the same knowledge-access problem appearing in a different task: useful assistance requires local system context.

By eleven, Hou is evaluating outputs and coordinating next steps. The component is complete with a Storybook link. The agents have identified and reverted the bad commit, and they have begun a postmortem and remediation exploration. His role has shifted from implementing each task to evaluating the work and deciding what happens next. He prepares to manage deployment of the fix while the agents handle loose ends and continue through lunch.

Agent icons connect to a daily timeline beside an “Evaluate, Coordinate, Review” card asking about gRPC migration, RFC readiness, and deployment timelines. The 11 AM review step is highlighted; later activities remain dim.Open full source frame
The timeline highlights evaluating, coordinating, and reviewing parallel work at 11 AM.
5:105:21
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:10 · section reference included

Remediation still ends with human review

After lunch, the broader remediation is ready for Hou’s attention. Hou reports that the gRPC library upgrade remediation touched 12 services and 20,000 lines of code. The agents produced tests and a writeup, and an engineering peer thanked him for getting the work across the line. Hou’s comparison is with what an organization might estimate as three engineering weeks of work, not a measured control project. The result was tested and almost ready for deployment, with a final round of human polish still required.

The benefit extends beyond doing several tasks at once: Hou recovers time that would otherwise disappear into the incident and its aftermath. Hou reports handling the three challenges in around half a day of active keyboard time. That is his account of hands-on effort, not total elapsed agent runtime or a general productivity benchmark. The complexity and urgency remain; what changes is how much work must pass serially through one engineer’s hands.

7:357:47
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:35 · section reference included

Mentor for the missing context

The working relationship resembles mentoring a junior engineer: provide the context and structure needed to do useful work, then evaluate the result. Both a new hire and an AI agent arrive without the organization’s history, its implicit assumptions, or years of experience with its particular systems. They may be able to implement a task in isolation while still missing why that implementation would be wrong for this codebase.

The analogy becomes less accurate when learning and retention enter the picture:

DimensionJunior engineerAI agent in Hou’s framing
Initial contextMust learn the organizationMust receive organizational context
Learning and executionSlower acquisition and executionRapid processing of supplied information
RetentionRetains and synthesizes experienceNeeds persistence across conversations
ConcurrencyLimited personal attentionCan support simultaneous tasks

Hou describes AI as perpetually junior because conversational experience does not automatically become durable knowledge. Product-supported persistence, such as Augment’s Memories, can carry guidance across conversations; it does not remove the need to supply and maintain that guidance. Engineers therefore become ongoing technical leads, responsible for structuring work and evaluating what agents produce.

That responsibility becomes more visible at team scale. Individual engineers can get strong results while the organization struggles to reproduce them. At Augment, colleagues asked how they could get the agent Hou had. The question exposes two separate challenges: spreading an individual’s effective practices and turning team productivity into a sustainable business advantage.

Hou cites a six-month new-hire ramp-up and four in five engineers naming context deficit as their biggest blocker, without identifying the studies or populations behind those figures. His diagnosis is that the knowledge problem predates AI: human teams already struggle to access what they need. Adding agents does not make scattered decisions, undocumented assumptions, or inaccessible project history disappear.

The proposed remedy is knowledge infrastructure: tools and systems that make organizational knowledge available as part of normal work. That does not require indiscriminately writing more documents, reorganizing every knowledge store, or rebuilding the company around an AI tool. The infrastructure should serve the team by making its existing knowledge usable.

8:529:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:52 · section reference included

Gather knowledge, learn the tools, then expand

Hou describes a three-step adoption sequence that starts with the organization’s existing material and expands as teams gain experience.

  1. Gather and connect knowledge. Inventory what already exists in Notion, Google Docs, GitHub, and other stores. Fill critical gaps around meetings and decisions, where reasoning and commitments can otherwise disappear. Hou uses Granola to capture meeting knowledge in most of his meetings outside engineering, producing tasks that can go into the tracker. Connect the resulting sources through MCP and native integrations so agents can use them during work.
  1. Build familiarity in both directions. Introduce the tools across teams and let engineers discover their strengths and weaknesses in real local tasks. At the same time, give the tools the organization’s coding patterns, architectural decisions, and business logic. Users learn how to guide the system; the system receives the context that makes its outputs useful.
  1. Expand successful practices. Give agents more complex tasks as demonstrated results build confidence. Share useful memories and task lists across teams so each engineer does not have to rediscover the same guidance. When colleagues asked for Hou’s agent, he shared his Memories file. The reusable asset was the accumulated guidance that helped the agent work effectively in their environment.

At the individual level, mentoring supplies missing structure and guidance. At the organizational level, shared knowledge infrastructure makes that guidance repeatable. The two activities reinforce each other: successful work reveals useful context, and making that context available improves the next person’s starting point.

12:2812:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:28 · section reference included

Return to the feature built during the talk

Hou returns to an extension development environment running a development build on Augment’s real codebase. Typing @ exposes the newly created personality, AI Engineer World Fair Augie. It even has a small icon, which he describes as rough. He asks for its favorite city, and the response enthusiastically chooses San Francisco, calling it the epicenter of the AI revolution.

The visible result is a selectable personality with an icon and behavior consistent with the request. This is a development demonstration, not a production deployment or a comprehensive validation of the implementation. Its significance is the overlap: the agent implemented a feature while Hou delivered the presentation. That opens a broader question about what teams can afford to explore when implementation requires less of their direct attention.

14:5315:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:53 · section reference included

Explore alternatives before committing to one

Once knowledge can move efficiently between people and agents, Hou sees an opportunity for parallel exploration. In a sequential design-build-test process, teams commit to decisions before they have much evidence about the resulting product. Each step narrows the remaining options. If prototypes cost less to build, teams can instead develop alternatives, iterate, put them in front of people, and converge using measured results.

Augment’s examples include a prototype VS Code fork kept available in case it becomes useful. The agent product itself began as a prototype, as did other features that engineers tried building with agents. Multiple approaches can coexist long enough to be tested before the team decides which one deserves the investment required for production. The diagram’s shorter time bands illustrate that proposed economic shift; they are not measured delivery guarantees.

Diagram titled “Unlocking parallel exploration: the economic transformation.” A sequential design, build, and test bar labeled “Week - Months” sits above parallel purple tracks labeled “Days - Weeks,” showing multiple approaches, simultaneous test, and rapid convergence.Open full source frame
Sequential development is contrasted with parallel approaches, simultaneous testing, and rapid convergence.

This changes when evidence enters a decision. Engineers often have to justify an architecture after investing in it, live with technical debt, or explain why a past choice became expensive. Leaders wish teams had more time for strategic work and fewer fires to extinguish. Testing divergent approaches early moves some of that learning ahead of commitment: designs, prototypes, and architectures become hypotheses that can be measured and validated.

Effective use of AI can make software creation more empirical. The opportunity is to compare working alternatives before committing resources, using tools that support both engineers’ guidance and the organization’s knowledge needs. Hou closes by announcing Remote Agents as available that week—a release announcement relative to the recording—and invites engineers to put parallel work into practice.

16:2416:39
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:24 · section reference included

Resources

From the talk

Updates since the talk

Transcript download
Full transcript

Loading transcript…