Dosu.dev on Using AI to Scale Engineering Knowledge and Reduce Developer Interruptions

Engineering teams rarely “run out of talent.” They run out of uninterrupted time. When it comes to execution, the delay isn’t due to a bug or the most complex architecture decision; it’s the steady drip of repeated questions, Slack pings, and context requests. Over weeks and months, those micro-interruptions compound into slower delivery, higher cognitive load, and a brittle knowledge culture where progress depends on a few people’s memories.

In a recent Software Plaza interview, Dosu.dev focuses on how knowledge that lives everywhere can be leveraged using AI-powered knowledge agents.

The hidden cost of repeated questions and Slack overload

Fast communication is both a boon and a bane for engineers. In a typical engineering channel, the same patterns repeat. Each question feels small, but answering it pulls a senior engineer out of flow, forces a context reload, and often triggers follow-up questions. Even when the answer already exists somewhere, such as an old PR thread, a changelog, a Confluence page, a Notion doc, or a GitHub README, it’s rarely discoverable when needed.

This is the heart of “tribal knowledge.” The knowledge isn’t truly absent; it’s just fragmented across systems and shaped by how differently every team structures it. That sprawl is why search alone often fails: even if you can find something, you can’t easily tell whether it’s still true, relevant to your environment, or the official answer.

Evolution of documentation

Documentation saw an evolution over time. While the ‘manual’ mode of updating documents supported the older model of business due to slower architectural evolution and fewer updates, as businesses focus on faster time to market, the processes within the organisation are also expected to keep up.

However, the problem with documentation is not a lack of information, but the way it is captured. Teams debate tradeoffs, make decisions, and encode those decisions in ephemeral channels. These communication channels then become the reference document that can take hours to search through to get answers.

Think of it this way: when you chat with a customer agent to track your order, vs. following the order page to track the shipment’s progress. Which of the two will be faster and revisited whenever necessary?

On the other hand, it is equally important to validate the ‘last updated’ field. A recently updated document does not imply it is accurate or relevant. The problem isn’t age, it’s traceability. Teams need a way to link answers back to sources so developers trust what they read.

Introducing AI-powered knowledge agents

A knowledge agent is not a regular chatbot with a large language model. The difference between a general chatbot and a knowledge agent, as the name suggests, is in the knowledge it possesses.

A generic chatbot has pre-trained knowledge without access to internal sources unless explicitly integrated.  A knowledge agent is already situated inside your engineering ecosystem. It ingests and continuously indexes the places your knowledge actually lives, GitHub, Slack, web docs, internal wikis, and then returns answers with citations back to those sources.

Dosu considers knowledge agents as “CI/CD for knowledge”. Instead of manually sweeping commits and conversations into docs, an agent can ingest changes as they happen and produce structured outputs, FAQs, onboarding guides, and change logs on demand.

Crucially, this shifts the model from “interrupt a human” to “ask the system.” The system can answer common questions directly in Slack, generate a first draft of a document tailored to a team’s sources, surface relevant snippets and links within the developer workflow, and reduce duplicate questions by making answers immediately available.

3 ways knowledge agents help with internal support

The role of knowledge agents is not to replace humans, but to leverage their time effectively to drive success in product development. Simply put, these agents reduce the additional workload on engineers, who inevitably spend time reviewing records and answering questions instead of doing actual work. Here are three ways in which knowledge agents can aid engineers. 

Improving support loop in Slack

With a knowledge agent in place, the need for a senior person to stop and respond to a query comes down significantly. The agents can respond with step-by-step instructions and even read citations from README sections, scripts, and past threads that justify the steps. The senior engineer only steps in if the question is novel or if the agent detects conflicting sources and escalates.

Better track of updates

The manual change summary is replaced with a changelog between dates, referencing merged PRs, release notes, and discussion context. Dosu believes that, as the knowledge agent generates these logs, there is a considerable reduction in human effort, as it eliminates the need to rummage through history and reduces interruptions to “who remembers why.”

Open-source maintainer triage at scale

Maintainers face a constant stream of issues, duplicates, and “is this on the roadmap?” questions, often from different time zones. Dosu’s value for maintainers showed up in automation around issue triage: tagging, deduplication, and answering common questions so maintainers can focus on higher-leverage work.

The implication for internal engineering teams is obvious: your “maintainers” are senior engineers and platform teams. If they spend their day triaging knowledge requests, they’re not building.

How to implement knowledge agents

Any organisation looking to implement agents inside production workflows must understand that trust is non-negotiable. To ensure all teams leverage the agents effectively, Dosu insists on having proper protocols that build credibility.

The first step would be to include citations. Every answer given by knowledge agents should point back to the exact source. This can be from documents, chat threads, or any other place where team members recorded their decisions. As a result, the answers become the source of truth, and developers can be assured they are not misinformed.

Another method is to use human-in-the-loop feedback. Users can rate answers as outdated, irrelevant, or incorrect, helping the system improve over time and align with team expectations. While this method doesn’t eliminate hallucinations, it contains them: the agent is constrained by your knowledge base and forced to show its work.

What the future could look like 

AI knowledge agents are not replacing documentation; they simply will make searching easier. Documentation may no longer be “one-size-fits-all.” One of the promising trends in documentation is dynamic docs tailored to the person’s role, environment, previous questions, and updates since last visit.

Instead of searching for answers in places where they work, such as Slack, GitHub, IDEs, and emerging standards like MCP servers, AI knowledge agents could let tools exchange context more directly.

When a team’s performance is affected by interruptions and tribal knowledge, the answer is not to prepare more documents, but to build a knowledge loop that continuously ingests decisions, makes answers discoverable in the flow of work, and reduces the need to tap humans for repeatable support.

That’s how engineering orgs scale: not by adding more people to answer the same questions, but by making the system itself capable of answering, with traceability, context, and minimal friction.