{ "video": { "id": "HvMyYLTfvhg", "title": "Building Agents Is Trivial Now, Context Is the Next Frontier — Jeff Ng, Unblocked", "duration": 802, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer" }, "analysis": { "video_id": "HvMyYLTfvhg", "title": "Building Agents Is Trivial Now, Context Is the Next Frontier — Jeff Ng, Unblocked", "one_liner": "Cloud primitives and frameworks have made deploying an agent nearly trivial, so the failure mode has shifted from infrastructure to missing organizational context — and Unblocked's Jeff Ng demos a Linear-triage agent that confidently recommends a fix that already caused an outage, until a \"context engine\" feeds it the Slack thread and postmortem it never saw.", "summary": "Jeff Ng, a founding engineer at Unblocked, argues that six months ago shipping an agent took a team a quarter because of infrastructure taxes — checkpointing and state persistence, sandboxing, observability — none of which make the agent smarter, and that Cloudflare/Vercel/AWS primitives plus frameworks have now absorbed that work, leaving model, instructions, tools, skills and sandbox location as the whole definition. He then shows the remaining failure: a Linear issue-enrichment agent, given only the ticket and the code, recommends re-enabling async dispatch to fix a 3–4 second time-to-first-character regression — a change a support engineer had deliberately disabled days earlier after it caused an outage. His diagnosis is that locally the human is the context layer, catching errors and supplying missing facts every turn, so removing the human from a background agent turns missing context into silent, compounding misinformation. The fix he proposes is a context engine that connects docs, code, tickets and conversations into a model of the organization and returns a reconciled, permission-scoped, synthesized understanding — not raw documents, which is what he says MCP alone gives you.", "key_points": [ "Six months ago building an agent took a team's effort and roughly a quarter, because an agent is not just models and tools but everything needed to run a production service.", "The infrastructure taxes he singles out: checkpoint/state persistence (agent runs are long-lived and stateful while infra is ephemeral — losing message history, tool calls and loop position means you can't resume, and restarting burns the tokens again, adds latency, and risks duplicating side effects); isolated sandboxes for agent-generated and third-party code (to prevent reads of environment secrets, unwanted network access, and taking down the shared host); and observability to answer \"where did this fail?\" across logs and traces from half a dozen systems.", "Cloud players (Cloudflare, Vercel, AWS) plus frameworks (he names Mastra, Vercel's SDK, and one transcribed as \"Flu\") have turned those into primitives; defining an agent is now just model, instructions/system prompt, tools, skills, and sandbox location — he says he was shocked how little code it took when he read the docs.", "Live demo 1: an issue-enrichment agent for Linear that fetches a ticket, classifies it as feature or bug, searches the code repository, and writes a plan of next steps back onto the ticket.", "The demo failure is real and specific: a colleague's ticket about agentic QA pipeline degradation where time to first character was 3–4 seconds instead of the expected hundreds of milliseconds. The agent recommends re-enabling async dispatch to parallelize the QE pipeline on a single machine — which had caused an outage days earlier and had been explicitly disabled by a support engineer.", "What the agent lacked was the Slack discussion held after the issue (the outage walk-through, the fix, next steps) and the resulting postmortem Linear ticket. As a background agent it would make that mistake silently, misinforming both teammates and other agents.", "His explanation for why this doesn't bite locally: the human is the context layer — asking questions, catching errors, supplying missing facts on every turn, knowing why the code is that way and what broke last time. The agent only has instructions, the tools and skills you gave it, the code, and the ticket.", "A context engine, as he defines it, supplies task-relevant information based on who you are and what matters, resolves conflicts across multiple data sets, respects the agent's access rules, and delivers a synthesized understanding the agent can act on rather than a list of documents it has to reason over itself.", "His case against just wiring up Slack/Linear/GitHub MCPs: MCP hands back raw results, floods the context window with irrelevant data, drives up context cost, and leaves the local agent to adjudicate ad hoc when the Linear MCP and the Slack MCP disagree.", "Demo 2 — same file, same engine, plus a call to the Unblocked context engine — surfaces the postmortem and the Slack conversation as a summary, and the recommendation flips from causing another outage to preventing one.", "Beyond ticket triage he claims the same layer helps coding agents (hydrating the agent plan saves context and tokens), code review (\"makes the PRs look as if they've been reviewed by an expert on your team\"), and surfacing correct answers for customer success and sales." ], "takeaways": [ "Stop building agent plumbing — checkpointing, sandboxes, tracing — yourself; take the primitives from Cloudflare/Vercel/AWS and a framework, since none of that plumbing improves agent capability anyway.", "Before deploying a background agent, ask what the human in the loop was silently supplying on every turn — the outage history, the decision that was reversed, the reason the code is the way it is — because that is what disappears when you remove them.", "Don't treat a pile of MCP connectors as a context strategy: they give access, not understanding, and they leave conflict resolution, ranking and permission scoping to the agent at inference time.", "Feed agents a reconciled, permission-scoped synthesis rather than raw retrieved documents — it both improves the answer and cuts context-window usage and token cost.", "Wire non-code sources — Slack decisions, documentation, postmortems, tickets — into the agent's context the same way you'd expect a new engineer to read them, especially for anything that runs unattended." ], "topics": [ "agents", "context-engineering", "mcp", "background-agents", "agent-infrastructure", "developer-tools", "rag", "observability" ], "tools": [ "Unblocked", "Cloudflare", "Vercel", "AWS", "Mastra (transcribed \"Maestra\")", "a framework transcribed as \"Flu\"", "Linear", "Slack", "GitHub", "MCP (Model Context Protocol)", "coding agents transcribed as \"cloud code or cortex\"" ], "quotes": [ { "text": "Everything I've mentioned here, none of this actually improves an agent's capabilities. They're all taxes one has to pay in order to get an agent out there to play the game.", "at": "02:43", "url": "https://www.youtube.com/watch?v=HvMyYLTfvhg&t=163s" }, { "text": "When a human is in the loop with the agent, we're there to catch the steer. Ultimately, we're there to babysit the agent.", "at": "07:42", "url": "https://www.youtube.com/watch?v=HvMyYLTfvhg&t=462s" }, { "text": "Scattered context comes in, grounded context comes out.", "at": "09:40", "url": "https://www.youtube.com/watch?v=HvMyYLTfvhg&t=580s" }, { "text": "MCP is great at access, but access isn't understanding.", "at": "09:57", "url": "https://www.youtube.com/watch?v=HvMyYLTfvhg&t=597s" }, { "text": "The gap isn't intelligence, it's context.", "at": "12:42", "url": "https://www.youtube.com/watch?v=HvMyYLTfvhg&t=762s" } ], "words": 2387 }, "summary_url": "/#HvMyYLTfvhg", "transcript": { "html": "/transcripts/HvMyYLTfvhg.html", "txt": "/transcripts/HvMyYLTfvhg.txt", "vtt": "/transcripts/HvMyYLTfvhg.vtt" } }