{ "video": { "id": "qdAkxLoYNI8", "title": "How to Generate Mergeable Code with a Context Engine — Peter Werry, Unblocked", "duration": 1116, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer" }, "analysis": { "video_id": "qdAkxLoYNI8", "title": "How to Generate Mergeable Code with a Context Engine — Peter Werry, Unblocked", "one_liner": "Unblocked's Peter Werry argues that coding agents are like new employees who reset their knowledge every task, and demos a \"context engine\" that pulls in Slack, PRs and Notion so Claude Code plans a fix in ~1 minute for under a dollar instead of ~2 minutes and more tokens.", "summary": "Werry frames the core problem as agents having access to code but not to intent, team conventions, past decisions or architecture rationale — the submerged part of the iceberg. He argues that dumping the codebase into a million-token window fails (distraction, wasted tokens) and that wikis fail because agents suffer \"satisfaction of search\": they find one plausible answer and stop. A context engine instead assembles task-specific organizational context, with sources shown so both humans and the agent can jump to the next thing; live demos cover codebase Q&A with a generated architecture diagram, a Slack bot, a Claude Code plan with vs. without Unblocked, a code-review agent, and a cloud agent that opened a PR correlating a fix back to the Slack thread that explained it.", "key_points": [ "Agents are \"expert software engineers who are new employees\": they rediscover your codebase, how you build, test and deploy on every single task, because knowledge resets per task.", "Uses a maturity-curve slide (attributed to \"Vim\"): autocomplete/Copilot in the GPT-3.5 days → Cursor → org wikis → MCP and skills; most people sit at stage 4–5, with \"software factories\" at stage 8 as where the puck is going.", "Attaching a wiki doesn't tell an agent where the information it needs is; agents hit \"satisfaction of search\" — the radiology failure mode of spotting one indicator on an X-ray and stopping before finding the others.", "Stuffing the whole codebase and architecture docs into context doesn't work even at a million tokens: there's more organizational context than fits, and it distracts the agent away from task-specific flow, wasting tokens and time.", "Live demo: asking about Unblocked's internal \"source mark engine\" produced an architecture diagram that doesn't exist anywhere — synthesized from how the code works today plus proposed future architecture — with sources shown as a trust and correction mechanism.", "Claude Code A/B on a plan to optimize the source mark calculator: with Unblocked, sub-dollar and about a minute; without, about two minutes and more cost, because the agent has to discover things — and it discovers the wrong things, so later execution runs on wrong assumptions and loops.", "The code-review agent mines pull request data to generate best practices that align agents to the codebase, and uses seniority/expertise as a signal to boost comments — it surfaced a prior comment from senior engineer Richie, who recognised it as something he'd said.", "A cloud agent debugged and fixed a precipitous drop in surfaced review issues, opening a PR whose description correlated the drop to a Claude model switch and linked back to the originating Slack conversation.", "Two open-source projects: a document query engine that ingests a repo's historical pull requests, synthesizes a schema from sampled documents and answers queries via agent chat; and an engineering social graph that maps code-review relationships into team clusters and shows where expert coverage has holes — the same signal used inside the context engine." ], "takeaways": [ "Stop treating a wiki or CLAUDE.md as solved context — measure whether the agent actually finds the right thing, since it will stop at the first plausible hit.", "Don't try to solve context by enlarging the window; assemble task-specific context per task instead, and treat distraction as a real cost alongside tokens.", "Compare agent runs with and without curated context on cost and wall-clock — Unblocked ships a free \"context engine simulator\" that runs a task both ways so you can see the delta before signing up.", "Always return sources with the context: it lets humans correct the knowledge base and lets the agent know where to jump next to elaborate.", "Judge context tooling on the compounding effect over long loops, not on the upfront cost of a short task." ], "topics": [ "context-engineering", "coding-agents", "code-review", "developer-tools", "organizational-knowledge", "mcp", "token-efficiency" ], "tools": [ "Unblocked", "Claude Code", "Cursor", "GitHub Copilot", "GPT-3.5", "Slack", "Notion", "GitHub", "MCP", "claude.md", "Sonar", "document query engine (open source)", "engineering social graph (open source)", "context engine simulator" ], "quotes": [ { "text": "Agents are like new employees. they reset their knowledge every time you start a new task.", "at": "01:54", "url": "https://www.youtube.com/watch?v=qdAkxLoYNI8&t=114s" }, { "text": "Access to information doesn't equal understanding.", "at": "04:13", "url": "https://www.youtube.com/watch?v=qdAkxLoYNI8&t=253s" }, { "text": "the real value of a context engine is not like the upfront cost on these short tasks. It's the compounding effect.", "at": "12:13", "url": "https://www.youtube.com/watch?v=qdAkxLoYNI8&t=733s" }, { "text": "50% fewer tokens, faster triage, better answers.", "at": "17:45", "url": "https://www.youtube.com/watch?v=qdAkxLoYNI8&t=1065s" } ], "words": 3429 }, "summary_url": "/#qdAkxLoYNI8", "transcript": { "html": "/transcripts/qdAkxLoYNI8.html", "txt": "/transcripts/qdAkxLoYNI8.txt", "vtt": "/transcripts/qdAkxLoYNI8.vtt" } }