Software Engineering Daily · softwareengineeringdaily.com

Organizational Context for AI Coding Agents with Dennis Pilarinos

·49 min·3 clips

Dennis used Claude Code to write a tool, Unblocked flagged Claude's insecure API key approach in code review, then Dennis asked Claude to fix it — and asks whether the pull request is becoming obsolete.

1. Software Engineering Daily host Kevin Ball interviews Dennis Pilarinos, founder and CEO of Unblocked, about building context engines for AI coding agents. 2. Dennis Pilarinos previously helped build Azure at Microsoft, worked at AWS, and co-founded BuddyBuild, a mobile CI platform acquired by Apple in 2017. 3. The episode's central argument is that as AI agents take on more code generation work, the primary bottleneck in software development has shifted from writing code to providing accurate organizational context about why systems were built the way they were. 4. Pilarinos defines context as 'decision-grade information' — the tribal knowledge that lets a person or agent understand architectural decisions, coding standards, and organizational constraints, rather than just the current state of the codebase. 5. Unblocked aggregates knowledge from Slack, Teams, GitHub (source code and pull request history), Jira, Notion, and production telemetry systems like Sentry and Datadog into a unified context engine. 6. A key problem Unblocked addresses is drift between sources of truth: source code may say one thing, a Slack thread another, a Jira ticket another, and the context engine must perform conflict resolution to avoid giving misleading answers. 7. Pilarinos describes Unblocked's 'killer scenario' as listening in shared Slack channels where developers ask infrastructure questions, and responding with high-quality answers within seconds — eliminating wait times for distributed or time-shifted teams. 8. The system supports multilingual teams: Portuguese-speaking developers can ask questions in Portuguese and receive responses in Portuguese, removing the barrier of asking in a second language. 9. Unblocked enriches organizational knowledge by running automated diffs on every pull request and building a background knowledge graph, because developer-written PR descriptions are typically too sparse to serve as historical decision records. 10. The platform uses a permissions-aware model that checks a user's access rights at query runtime, not at indexing time, so revoking a user's access to a document immediately stops Unblocked from surfacing that document's content. 11. Pilarinos demonstrates this with a 'Project Tantalus' scenario: a developer asks about a release date, gets an answer while they have Google Docs access, then immediately receives 'I don't know' after access is revoked — even though the data exists in the knowledge base. 12. Identity resolution across disparate systems — GitHub, Slack, Notion — is described as one of the harder infrastructure problems, because the same person may have different identities in each system and access controls must be correctly mapped. 13. The underlying architecture is described as a hybrid RAG system with agentic query expansion, using both lexical and semantic search, with observability tooling so agents can explain how they derived their answers. 14. Pilarinos quotes a claim circulating among developers: big banks have created more technical debt in the last six months from AI-generated code than in their entire institutional history — using it to illustrate the scale of the code review bottleneck problem. 15. The open-source community backlash against AI-generated pull requests is cited as evidence that the review burden from agent-written code is already causing friction for maintainers. 16. Unblocked's context-aware code review tool flagged a Claude Code security mistake in a real internal project — suggesting storing an API key in an environment variable — and recommended encrypting it in vault instead, which Pilarinos then asked Claude to implement directly from the review comment. 17. Pilarinos raises the question of whether the traditional pull request becomes obsolete when agents write code, context engines evaluate it against organizational standards, and automated tests validate it — potentially closing the loop without human code review. 18. The conversation style is technically substantive and collaborative, with host K-Ball frequently building on Pilarinos's points and contributing observations from his own engineering leadership experience. 19. Software engineers and engineering managers at organizations adopting AI coding agents will find the most practical value here. 20. Listeners looking for introductory coverage of AI in software development or business-level AI strategy discussion would find this conversation too technically granular.

As heard by us

A grounded look at why AI coding agents need organizational context, not just better code generation.

Software Engineering Daily uses Dennis Pilarinos's work on Unblocked to get at a practical problem in AI-assisted development: code is no longer the only bottleneck.

Read the full review in PlayNext →

Why you'd press play

Unblocked CEO Dennis Pilarinos on why organizational context has become the binding constraint on AI coding agents.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Software Engineering Daily