{ "video": { "id": "XqF-IFHBCkM", "title": "Redesigning How Software Gets Built With AI Agents — Sonar & McKinsey Panel", "duration": 1152, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "XqF-IFHBCkM", "title": "Redesigning How Software Gets Built With AI Agents — Sonar & McKinsey Panel", "one_liner": "Teams get to a fully autonomous PR pipeline one trust step at a time. Users start with agent reviews and CI failure analysis, then let the agent block PRs, then try fixes, then turn on auto-fix, then auto-approve non-critical changes, and finally auto-merge, each step unlocked by seeing the agent's precision.", "summary": "McKinsey and Sonar found that 80-90% of companies say they have deployed AI or agents, but fewer than a third have scaled them or see business impact. Most are at 'Horizon 2' (human-guided and validated agents) and very few are at 'Horizon 3' (a fully automated end-to-end workflow with human oversight). The panel names three factors that make scaling work: redesigning the process workflow, tooling, and, most often overlooked, people and the operating model, because the boundaries between roles are blurring. Sonar CEO Tariq and Ali (formerly CEO of Gitarr, now part of Sonar) add practical lessons: be clear about what you measure, run lighthouse teams, use platform teams and baseline metrics, make implicit knowledge explicit, and build trust step by step toward autonomous PRs.", "key_points": [ "McKinsey research: 80-90% of companies say they have deployed AI or agents in some form, but fewer than a third have scaled or see business impact. Many are stuck in POCs and small environments.", "Maturity scale: most companies are at Horizon 2 (human-guided and validated agents executing tasks across the SDLC). Few are at Horizon 3 (a fully automated end-to-end workflow with human oversight).", "Three scaling factors: redesign the end-to-end workflow, have good tooling (agents, harnesses, context layers), and remodel the operating model around blurring roles (engineer, PM, designer). Organisations that did all three see the 'promised land of 10x productivity'.", "Sonar, which started as an open-source project 18 years ago, organised its own engineering around its 'guide verify solve' model. It rejects 'token maxing' as a goal and focuses on generating real value. It used lighthouse teams to model practices such as PRD agents that generate the evals used for AI output verification.", "Tariq's surprise: lighthouse teams and role modelling got the other teams only part of the way. Getting people to think differently and understand why they use agents is 'one of the critical missing links'.", "Ali (formerly ran the developer platform org at Uber, with platform roles at Google and Facebook before that) argues large companies have a structural advantage: an existing platform team. That team needs baseline metrics before adding agents: speed to production, throughput, quality, how often work goes back to the start of the workflow, and engineer sentiment. It should automate whole workflows rather than single steps, and use agents to remove grunt work so developers stay in flow.", "Tariq's view of the future: no fully hands-off world; 'orchestrated' is the key word. Implicit company knowledge (preferred code styles, dependencies) must be made explicit. Critical thinking matters because AI gives 'very high conviction plausible output'. This is 'the era of the generalist'. Resistance to smaller teams usually comes down to siloed role identities like front-end versus back-end.", "Trust ladder Ali sees at Gitarr: reviews plus CI failure analysis → blocking rules on issues the agent finds → trying automated fixes → turning on auto-fix that iterates until the build is green → rules to auto-approve non-critical code areas or change types → auto-merge, including fixing merge conflicts. He expects more teams, including in large companies, along this path within the next year." ], "takeaways": [ "Before adding agents, set up baseline developer-productivity telemetry (speed to production, throughput, quality, rework, sentiment) so you can measure, iterate and adjust.", "Measure real value, not token usage, and automate whole workflows end to end rather than adding a tool to one SDLC stage.", "Don't assume lighthouse teams will spread practice on their own. Invest in helping other teams understand why they are using agents, and restructure roles and teams around generalists and smaller units.", "Write down the implicit knowledge in your team's heads (conventions, preferred dependencies, what success looks like) so agents can use it.", "Grant agents autonomy in stages: review, then block, then fix, then auto-fix, then auto-approve low-risk areas, then auto-merge. Move up a step only once you have seen precise output and have rules and guardrails in place." ], "topics": [ "agents", "sdlc", "operating-model", "developer-productivity", "code-review", "trust", "platform-engineering", "org-design" ], "tools": [ "Sonar", "McKinsey", "Gitarr (also transcribed as Gitai/Guitar)", "Uber", "Google", "Facebook", "Anthropic (transcribed as 'Tropic', re fable models)", "Twitter" ], "quotes": [ { "text": "The objective was never for us token maxing or anything like this.", "at": "04:24", "url": "https://www.youtube.com/watch?v=XqF-IFHBCkM&t=264s" }, { "text": "AI always gives you very high conviction plausible output", "at": "13:17", "url": "https://www.youtube.com/watch?v=XqF-IFHBCkM&t=797s" }, { "text": "I think this is the era of the generalist", "at": "14:28", "url": "https://www.youtube.com/watch?v=XqF-IFHBCkM&t=868s" }, { "text": "now what you've got is fully autonomous end-to-end PR creation through merge is fully auto- automated. But it doesn't come immediately.", "at": "18:08", "url": "https://www.youtube.com/watch?v=XqF-IFHBCkM&t=1088s" } ], "words": 4131 }, "summary_url": "/#XqF-IFHBCkM", "transcript": { "html": "/transcripts/XqF-IFHBCkM.html", "txt": "/transcripts/XqF-IFHBCkM.txt", "vtt": "/transcripts/XqF-IFHBCkM.vtt" } }