{
  "video": {
    "id": "OL7kfezynJM",
    "title": "Multiplayer agentic engineering — Arjun Singh, Superconductor",
    "duration": 1124,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "OL7kfezynJM",
    "title": "Multiplayer agentic engineering — Arjun Singh, Superconductor",
    "one_liner": "Arjun Singh of Superconductor lays out six lessons from a year of running agents across a whole team — be model and harness agnostic, put the same agent session into every human interface, and benchmark agents on your own code base rather than trusting SWE-bench.",
    "summary": "Singh argues that most agentic-engineering talk centres the agents while ignoring the people, and walks through the friction points his team (the ex-GradeScope crew, now building Superconductor/Superagent) hit while pushing agents into every workflow. The lessons: stay model- and harness-agnostic; make one agent session reachable from Slack, the app and GitHub; make agent work visible via artifacts like screenshots; auto-ingest external signals (meetings, Sentry, email) into evaluable code; move everything into isolated cloud sandboxes with network egress control; and benchmark agents on your own repo. He demos a meeting bot that listened to a 4-hour expo-booth Google Meet, turned an attendee's idea into a ticket and a working change unprompted, and shows internal quality-vs-cost and quality-vs-time charts across harnesses on their Ruby on Rails code base.",
    "key_points": [
      "Be model and harness agnostic because 'the best model and harness can change weekly' — either a new one ships or the best one gets taken away; he cites Fable arriving, becoming their default for a few days, then going away and them switching back to Codex with no disruption.",
      "Open-weight models are now viable: they've been 'really happy with GLM 5.2', it's much cheaper, and more and more of their work is going through it.",
      "Slack bots (Claude's, Codex's, theirs) aren't enough — they move an agent from trapped on one laptop to trapped in Slack. The goal is the *same* agent session accessible from Slack, the desktop/mobile app and GitHub, so it never forgets context across interfaces; the app view shows who created, commented on and saw a session, which matters when non-technical people trigger work.",
      "Their meeting bot (invited to Google Meet/Zoom/Teams) ran all day at their expo booth on a 4-hour call, links to existing work rather than duplicating it, and autonomously turned an attendee's suggestion — that agents should have clear acceptance criteria before declaring done — into a ticket plus a change adding two acceptance-criteria fields to their ticket form, then produced a screenshot on request. He says every onboarding/customer/team call now yields dozens of prototyped ideas and at least a few shippable PRs with minimal intervention.",
      "Isolated cloud environments do three things: kill 'lid anxiety' (his motivation was not wanting to be tethered to a laptop with a 6-month-old at home), enforce least privilege so an agent can't find a stray token, think it's on staging and wipe production, and add a configurable network sandbox that prompts for per-ticket or per-project egress access so code and secrets can't be exfiltrated in YOLO mode.",
      "Sandboxes are what let support and growth people trigger real work — they have no dev environment, so they ask in Slack, the fix lands with screenshots, an engineer reviews and merges, skipping Linear triage and a PM.",
      "Benchmark on your own repo by picking PRs that represent great engineering work (agent, human or hybrid) and running candidate agents to get quality-vs-cost and quality-vs-time; public benchmarks may not transfer — 'swe bench is all in Python, we're Ruby on Rails.' Their results: Anthropic agents consistently better but not faster and clearly the most expensive; Codex and Cursor fast and quite good; open models improving but slow. Hard data matched their vibe check and they switched their default to Codex.",
      "Scale numbers for a relatively small team: ~99.9% of PRs heavily agent-generated but everything human-reviewed; 1.5 billion tokens in the past month; 3,300 Claude Code runs at $10,000/day in token-equivalent cost (covered by plans); Codex with 4x the sessions and cheaper overall. Next step is using own-codebase benchmarks to route tasks automatically, since no third party knows what works on your repo."
    ],
    "takeaways": [
      "Get your code base and agents running in an isolated cloud sandbox with a configurable network allowlist — he recommends doing it even with a competitor's tooling, and says you can just have Claude Code or Codex set it up for you now that agents are good enough.",
      "Wire agents into the interfaces your team already uses (Slack, GitHub, app, mobile) with session continuity, so nobody has to context-switch or copy context between tools; roll your own if you don't buy a product, because the friction otherwise is too high.",
      "Build a benchmark from your own repo's exemplary PRs and score candidate agents on quality vs cost vs time, instead of extrapolating from SWE-bench or Terminal-Bench.",
      "Stay model- and harness-agnostic so a new release, a withdrawn model, or a cheap open-weight option like GLM 5.2 can be tried in minutes rather than becoming a two-hour experiment you keep postponing.",
      "Auto-ingest external signals — meetings, Sentry, bug reports, feature-request emails — into prioritised, already-started tickets rather than relying on humans to shuttle 'solve ticket 6,000' into an agent."
    ],
    "topics": [
      "agents",
      "coding-agents",
      "team-collaboration",
      "sandboxing",
      "benchmarking",
      "model-routing",
      "agent-security",
      "developer-workflow"
    ],
    "tools": [
      "Superconductor",
      "Superagent",
      "GradeScope",
      "Claude Code",
      "Codex",
      "Cursor",
      "GLM 5.2",
      "Fable",
      "MiniMax",
      "Kimi K2",
      "Slack",
      "GitHub",
      "Google Meet",
      "Zoom",
      "Microsoft Teams",
      "Sentry",
      "Notion",
      "Linear",
      "MCP",
      "SWE-bench",
      "Terminal-Bench"
    ],
    "quotes": [
      {
        "text": "the incentives of the people selling you tokens aren't really aligned with yours. You're here for a reason... They want to sell you more tokens. And you might be happy to pay for as many tokens as it takes, but you don't want to pay for more than that.",
        "at": "02:17",
        "url": "https://www.youtube.com/watch?v=OL7kfezynJM&t=137s"
      },
      {
        "text": "Cuz now we've taken it from trapped on somebody's laptop to kind of trapped in Slack.",
        "at": "03:25",
        "url": "https://www.youtube.com/watch?v=OL7kfezynJM&t=205s"
      },
      {
        "text": "the first one is to eliminate what some people are calling lid anxiety. You want to be able to close your laptop.",
        "at": "09:13",
        "url": "https://www.youtube.com/watch?v=OL7kfezynJM&t=553s"
      },
      {
        "text": "Like swe bench is all in Python, we're Ruby on Rails. It is not the case that the benchmarks are identical for them.",
        "at": "13:55",
        "url": "https://www.youtube.com/watch?v=OL7kfezynJM&t=835s"
      }
    ],
    "words": 4533
  },
  "summary_url": "/#OL7kfezynJM",
  "transcript": {
    "html": "/transcripts/OL7kfezynJM.html",
    "txt": "/transcripts/OL7kfezynJM.txt",
    "vtt": "/transcripts/OL7kfezynJM.vtt"
  }
}