AI agents work best when engineers give them outcomes, context, constraints, and evaluation instead of handing them isolated tickets.
2
Augment Code uses agents to run parallel workstreams across implementation, incident response, documentation, and team communication.
3
Organizations need shared knowledge infrastructure before individual productivity gains from AI can spread across teams.
Summary
Eric Hou describes how Augment Code uses AI agents while building production software. His central comparison is that agents behave like perpetually junior engineers: they lack system and organizational context, forget between conversations, and need guidance, but they can work quickly and on several tasks at once. He recounts a Tuesday when agents helped him scope a design-system component, investigate a staging incident, support a new hire, revert a bad commit, and remediate a gRPC upgrade affecting 12 services and 20,000 lines of code. His role shifted toward evaluation and coordination. Hou then explains why individual success is difficult to scale. Teams need knowledge infrastructure that connects documentation, code, decisions, meetings, and task systems. He proposes gathering existing knowledge, helping people and tools learn their working context, and spreading successful memories and task patterns. With this foundation, teams can prototype competing approaches cheaply and use real measurements to make software decisions earlier.
Agents should receive outcomes and boundaries instead of detailed implementation instructions
Hou says his first effective use of an agent was to provide the outcome, context, and constraints, then let it work through the codebase. At 9:00 a.m. in his Tuesday example, he asks an agent to scope a design-system component and build an RFC according to the team's architectural patterns. While he takes a coffee break, the agent explores the repository and produces a mostly completed RFC. The point is to create enough structure for the agent to operate while leaving the engineer responsible for judging the result. Hou describes this as scaffolding and providing context, rather than micromanaging every implementation step.
Parallel agents let one engineer keep the critical path while other work continues
When a staging emergency interrupts the design-system work, Hou does not abandon the component. He hands it to an agent and starts separate agents on log analysis and git bisect. An Augment Slackbot handles communication with teams blocked from deploying. Hou stays focused on the on-call problem while several workstreams proceed in the background. By 11:00, the component is complete, the bad commit has been found and reverted, and agents have begun a postmortem and remediation work. His job has shifted from doing every implementation task to evaluating results and coordinating the next steps.
The agent's speed does not remove the need for human judgment
Hou says the agents completed a remediation for a gRPC library upgrade that touched 12 services and 20,000 lines of code. The work included tests and a written explanation, and an effort that a normal organization might estimate at three weeks was nearly ready for deployment after half a day of Hou's active keyboard time. He still describes a final round of human polish as necessary. The value comes from recovering time under pressure, rather than from eliminating engineering responsibility. Humans decide whether the work fits the system and whether it is ready to ship.
AI resembles a junior engineer in context, but it learns and forgets differently
Hou identifies a shared context or knowledge gap between AI and new engineers. Both start without the organizational and system knowledge needed to implement safely. A junior engineer learns slowly but retains and synthesizes what they learn. AI can process the same information and implement a request in minutes or seconds, yet it forgets between conversations. Hou therefore calls AI a perpetually junior engineer that can work on multiple tasks at once. Engineers need to act as perpetual tech leads, giving agents the mentoring and structure that junior engineers also require.
Individual productivity gains stall when an organization cannot supply shared context
Hou says teams often struggle to reproduce the results of individual engineers. At Augment, people asked how they could get the same agent experience he had. He connects this problem to a wider organizational weakness: new hires may take six months to ramp up, and engineers commonly identify context deficits as a major blocker. The same missing knowledge limits AI. Hou argues that scaling agents requires knowledge infrastructure that connects the organization's code, documentation, decisions, and working practices. Adding more documents alone does not solve the problem if the information remains disconnected from the tools and workflows where agents operate.
Knowledge infrastructure starts with existing sources and captured decisions
Hou's first adoption step is knowledge gathering. Teams should map the sources they already have, including Notion, Google Docs, and GitHub, then identify important gaps around meetings and decisions. He describes using Granola AI to record meetings and produce tasks that can go into a task tracker. He also recommends integrating data sources through MCP and Augment's native integrations. The goal is to build an initial knowledge base from information that already exists, while capturing decisions and conversations that would otherwise disappear. This gives agents more of the context they need to work within real organizational constraints.
Teams improve when people and tools learn the organization's working patterns
The second adoption step is gaining familiarity with the tools. Hou says this applies both to engineers learning the tools and to the tools learning the organization. Teams should try agents in their own context and learn where they work well or poorly. Over time, they can teach the platform about coding patterns, architectural decisions, and business logic. The third step is to expand successful patterns and assign more complex tasks as trust grows. Hou describes sharing Augment memories and task lists across teams so that one engineer's successful approach can spread instead of being recreated from scratch.
Shared context makes parallel prototyping a way to test decisions earlier
Hou argues that knowledge infrastructure changes the economics of software development. When information transfer becomes fast and scalable, teams can prototype several approaches, test them, and converge on decisions using real metrics and user feedback. Augment keeps prototypes available, including a possible VS Code fork, and the agents product itself began as a prototype. Engineers can try several designs or architectures before the organization commits production resources. Hou says this turns decisions that were once argued about or regretted later into hypotheses that can be measured early. He concludes that effective AI use can make software creation more like a science because teams can test alternatives before locking them in.
"To make the most use out of AI, we need to work with it as we would work with junior engineers. Not assigning tickets, but mentoring."09:00
Who should watch
You are an engineer losing planned work to incidents, interruptions, and support requests, and want a concrete way to run those tasks in parallel.
You lead an engineering team that has seen strong individual results from coding agents but cannot spread those practices across the organization.
You are building internal AI workflows and need to connect documentation, code, meetings, decisions, and task systems before assigning agents larger responsibilities.
