Owen Jennings
Business lead at Block (Square, Cash App, Afterpay); architect of Block’s AI transformation including the Goose agent harness, the ~40% workforce reduction, and the “proprietary signal + iteration loop” moat thesis.
Last updated: 2026-04-16
Overview
Owen Jennings joined Cash App in 2016 during its monetization phase and helped grow it to ~60% of Block’s gross profit. As business lead overseeing product operations and customer support across Square, Cash App, and Afterpay, he was the executive who orchestrated Block’s AI transformation and the roughly 40% reduction in force in early 2026 — the first major public company to make a decision this explicit about AI displacing headcount.
His framing: this wasn’t cost-cutting; it was redesigning the company around what software development actually looks like now. The cuts were heaviest on the development side, light on outbound sales and account management — the opposite of typical overhang-driven RIFs.
The December 2024 Inflection Point
Block launched Goose (their agent harness) in early 2024. Throughout 2024 and into 2025 there was meaningful progress, but tools were “pretty good at writing code especially for new ventures and green space.”
Then — late November / first week of December — a binary change. Owen names Opus 4.6 + Codex 5.3 as the specific models. The tools became capable of working with existing complex codebases overnight. The correlation between headcount and output that had held for decades broke in that week.
“We’re not writing code by hand anymore. That’s over. That’s done.”
This aligns directly with Karpathy’s December 2024 inflection report (he stopped writing code by hand at the same time). Block spent Q1 as an executive team with Jack Dorsey working through the implications, then executed.
Goose and the Agent Infrastructure
Goose — Block’s open-source agent harness, launched early 2024. Owen believes it was the first agent harness of its kind. Key properties:
- Model-agnostic: supports ~120 models; operator swaps between Anthropic, OpenAI, open-source depending on task
- Foundation for all Block automations: Moneybot (Cash App), ManagerBot (Square), and internal automations all route through Goose
- Became the substrate for Block’s “agentic operating system”
G2 — internal agentic operating system. Anyone at Block can automate any deterministic workflow through G2. The platform abstraction above Goose that makes automation accessible to non-engineers.
Builderbot — internal tool analogous to Claude Code but deeply integrated with Block’s infrastructure. Key capability: autonomously merging PRs and building features to completion. Current pattern:
- Complex features built to 100% autonomously (happens)
- More common: built to 85–90%, then a human with deep context finishes the final 10%
- Time from idea → 100K+ customers has been “compressed massively since December”
Org Structure After the RIF
The new structure is built around agents as the execution layer:
| Old | New |
|---|---|
| Feature teams of ~14 (8 backend, 4 client, 1 PM, 1 designer) | Squads of 1–6 people |
| Functional silos (“I’m on the banking team forever”) | Fluid squads: work a few cycles on one product, then move |
| Hierarchical layers (~5–7) | 50–60% fewer layers; product org has 2–3 layers max |
| Linear workflow: write PR → submit → review → fix | 14 agents building PRs simultaneously; human context-switches between them |
| Back-to-back meetings | 70–80% meeting reduction; time returned to building |
| Weekly sprint cadence | Weekly all-hands with Jack; async, agent-mediated work |
All designers are shipping PRs. All PMs are shipping PRs. That’s no longer noteworthy.
The new work pattern: “In the background, 10–20 agents are doing a bunch of stuff, and I have to check in, nudge, change, commit.” Not a linear workflow — a supervision + editing loop.
Compliance exception: the compliance team was not significantly touched. Regulatory relationships require human accountability; the risk of getting it wrong is asymmetric.
Generative UI Thesis
Block’s product bet: the era of static, uniform UI is ending. Within 6 months of the interview (mid-2026), everyone’s app will look different.
- Level 1 (achievable with personalization): different content surfaces based on usage patterns. Not interesting.
- Level 2 (what Moneybot does now): on-the-fly chart and visualization generation in response to queries. “How have I been spending?” → the model generates the visualization at runtime, not from static code.
- Level 3 (ManagerBot on Square): generate entire mini-apps on demand. A multi-location restaurant owner asks “build me a scheduling app with WhatsApp notifications for my employees” — ManagerBot creates that app, and its UI is not in the source code of the shipped application.
The QA problem: non-deterministic output for tens of millions of customers is a genuine unsolved engineering challenge. Block is actively working on it.
Proactive intelligence: don’t ask customers to prompt; surface insights to them. Owen’s finding: customers (especially around money) don’t know the right prompts. The value comes from the system noticing things and surfacing them proactively, not from building better prompt interfaces.
The Moat Thesis: Proprietary Signal + Iteration Loop
Owen’s answer to “what makes a company defensible in the AI era”:
Near/medium term moats that still exist:
- Distribution and network effects (nobody’s vibe-coding 50–60M monthly actives)
- Regulatory licenses and posture
- Hardware (can’t vibe-code Square hardware)
Long-term moat (the only durable one):
“The biggest moat is going to be which companies understand something that’s super hard for other people to understand. And if your answer to that is ‘I don’t know,’ then you maybe could get vibe coded away.”
The flywheel:
- Proprietary signal — rich data and deep insight about a domain (for Block: how sellers and buyers participate in the economy)
- World model — internal markdown of who you are, values, metrics, what you care about
- Iteration tool — Builderbot / Claude Code / equivalent
- Tight feedback loop — run the cycle hundreds or thousands of times a day; humans as editors, not builders
This is the AI-era equivalent of compound interest: the company that understands its domain deepest and can iterate fastest builds an insurmountable lead. Companies without a clear proprietary signal are vulnerable to being “vibe coded away” by faster-moving competitors.
Jevons paradox: Owen’s explicit view — fewer engineers per product, but more total products, more companies, more sectors getting software for the first time. Doesn’t necessarily mean fewer engineers in aggregate.
Connections
- product-trio-agentic-era — Block is the most explicit practitioner data point; Owen confirms: all designers/PMs shipping PRs, small squads, editors not builders, supervision loop replaces linear workflow
- agent-first-software — generative UI thesis; “vibe coded away” framing; Moneybot/ManagerBot as agent-first product layer; proprietary signal moat
- agentic-engineering — Goose as a real-world four-layer implementation; G2 as orchestration layer; Builderbot as capability layer with guardrailed autonomous PR merging
- auto-research — “14 agents building PRs simultaneously; I context-switch” is exactly the SETI@home distributed parallel pattern Karpathy described
- thin-harness-fat-skills — Goose is the thin harness (~120 model-agnostic); Moneybot/ManagerBot are the skills built on top
- andrej-karpathy — both name December 2024 as the binary inflection point for coding with complex existing codebases
- jensen-huang — Jevons paradox: more products and sectors, not fewer engineers total; five-layer stack thinking maps to Block’s agent substrate work
Sources
- How to Reorg After AI Changes Everything — Owen Jennings, Block (YouTube) — added 2026-04-16