Dennis Pilarinos from Unblocked: context for AI coding agents
Unblocked says organizational context makes coding agents 40–60% faster while cutting token use by 30–50%—because every newly spawned agent otherwise arrives like an expert developer on their first day at the company.
Dennis Pilarinos is the founder and CEO of Unblocked, a context layer that connects information across the systems a software team uses and makes it available to people and coding agents.
Recorded at LeadDev in London, this conversation traces the path from hand-maintained context files to MCP servers and organization-wide knowledge graphs. Dennis explains why agents stop searching too early, how conflicting and stale information is reconciled, and why identity and permissions become foundational when context spans GitHub, Slack, Notion and other systems.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
How has Unblocked changed as coding agents have become more popular?
Unblocked began as a context layer where people could ask questions about their organization. As agents started writing more of the code, they needed access to the same information. We have focused on serving that context to agentic workflows, and we are seeing tasks completed 40–60% faster with 30–50% fewer tokens for a given job.
What context is a newly spawned coding agent missing?
It does not know why the codebase works as it does, how decisions get made or what the team learned through pull-request reviews and postmortems. Every new terminal is like a developer's first day at the company. The agent may be highly capable, but it wastes time grepping the repository and calling tools because it has none of the organizational experience that makes a long-serving engineer effective.
How do teams mature from local context files to a real context layer?
The first stage is often a personal collection of files, perhaps in Obsidian, that tells an agent how to behave. Teams then merge those files into a shared but static source that immediately begins going stale. Exposing business systems through MCP is another step forward, but agents still struggle to reason across them. The final stage pulls the information together into a knowledge graph and provides reconciled answers across the organization.
Why isn’t giving an agent access to several MCP servers enough?
Agents exhibit 'satisfaction of search,' a term from radiology: once they find one plausible result, they stop looking for other evidence. An agent may search Notion, see something that appears correct and never inspect Slack for information that changes the conclusion. MCP provides access and reduces some content rot, but access alone does not create an exhaustive, reconciled view of the truth.
What makes context more important than raw model intelligence?
Imagine two engineers with ten years of experience: one has spent those ten years at your company, while the other has just arrived. The first will usually make the better architectural decision because they understand the local intent and history. Coding agents are expert strangers. Intelligence is not the principal gap; context is.
Can Unblocked capture knowledge that has never been written down?
That remains harder because the information has to be transcribed in some form before software can use it. Even with written sources, the difficult work is deciding what is true when documents conflict or one supersedes another. The context layer must reason about those differences rather than returning the first matching passage.
Why are identity and permissions difficult in an organization-wide context layer?
The same person has different identities in GitHub, Slack, Notion, Confluence and other systems. To answer a question safely, Unblocked has to resolve those accounts to one person and determine their permissions at runtime. Someone should receive information from a spreadsheet only if their corresponding identity is allowed to access it. When that works, the complexity is invisible and the user simply gets a fast, authorized answer.
Do teams need to clean up their documentation before using Unblocked?
No. Asking an organization to rewrite its workflow, delete obsolete Slack messages or clean every Confluence page creates a large upfront cost for a benefit it may receive much later. Unblocked connects to the sources as they are, then cleans and enriches the information while building the knowledge graph. The team keeps working normally and the agent receives better answers.
What are you most excited about now?
We have been talking about context for years, and every idea has its time. It is exciting to see teams recognize that they have this problem and that solving it produces a meaningful improvement. The shift from humans writing all the code to agents participating in the work has made the need for an organizational context layer much easier to see.