Jellyfish

Spent the last week at Jellyfish researching what autonomy in coding agents actually looks like in practice, specifically for Claude We found out that most sessions are short, but once you zoom into the top 0.1%, the picture changes and a lot Three things jumped out: 1. Productivity scales steeply with duration: Going from sub-P90 turns to P99.9+ turns, commits per turn go up roughly 20x and merged PRs about 11x. The work mix barely shifts — slightly more features, slightly less refactor — so these long runs look like generalists doing way more, not specialists doing something different. 2. Over 80% of long-running turns happen with the human barely involved. And with new models like Opus 4.7 pushing how long agents can stay on task, coding agents are entering another era of scale and volume. 3. Output concentrates in those low-supervision long runs. Average net lines of code lands around 11,800 per turn, vs. ~1,000 when supervision is high for the same duration. About 11x. My guess for 2026-2027: the bottleneck shifts away from model coherence and toward CI queues, review capacity, and whatever infra has to sit underneath fleets of these agents running in the background. Having a surface area for agents to interact 24/7 will become business critical. Link in comments to the complete details. #AI #Agents #Claude #Autonomy

“The most aggressive adopters of AI are beginning to pull away from more hesitant competitors,” said Andrew Lau, CEO and co-founder of Jellyfish. “The role of the engineer is transforming as a result. Data is the difference between being overwhelmed by this wave of change and marshalling the new tools effectively to unlock new levels of AI potential. Organizations need hard data to drive measurable impact using AI coding tools and communicate their ROI to the C-suite and the board beyond.” Learn more about navigating the next wave of AI transformation here: https://lnkd.in/gvYe5t9t #SEMR26 #StateOfEngineeringManagement

Had another great conversation this week with Alistair Barr at Business Insider, this time about the tokenmaxxing phenomenon and what Jellyfish data has to say about the limits of raw AI consumption. https://lnkd.in/eruX9rHH Top Claude Code users are consuming about 10x the tokens of the median developer, but only producing about 2x the output. Tokens are like rocket fuel: incredibly powerful, but you need increasing amounts to go incrementally faster. There's a broader shift happening across the industry right now. We're moving past "adopt AI at any cost" to "adopt AI, but show the receipts." Companies are willing to spend here, but increasingly need to show they're spending responsibly and having an impact. Every engineering leader knows that the real job isn't just proving what's possible – it's building and scaling an entire machine. When it comes to AI usage, you'll get way more leverage by getting everyone into the middle of the adoption curve than by pushing a small group into the stratosphere.

The 2026 State of Engineering Management Report is here 🎉 For the seventh consecutive year, Jellyfish surveyed 635+ engineering leaders and practitioners worldwide to understand how the discipline is evolving. This year, the story is clear: engineering's impact has never been greater – and AI is the reason. Explore the report's findings to answer the questions engineering leaders everywhere are asking, like: - Is everyone ahead of me on their AI adoption journey? - What AI tools should my team actually be using? - What kind of productivity lift can we really expect from AI? - How do we handle the ballooning cost of AI tools? https://lnkd.in/g5UVVrH8 #SEMR26 #StateOfEngineeringManagement

In prep for BosTech Week, join our "From Zero to Extension: Build Your Own Agentic Workspace" hackathon co-hosted with Nimbalyst at Pillar VC In 90 minutes, you'll plan, build, and deploy your own extension for an agentic workspace powered by Claude Code and Codex. Not a demo. Not a tutorial. Something you could actually use at work. Details below: 📍 PillarVC, Boston 🗓️ Thursday, May 21 🍕 Pizza and drinks provided Whether you're a developer or a founder, if you're thinking about how AI is reshaping engineering, this one's worth your Thursday night: https://luma.com/xxf112kf #BostonTech #BosTechWeek #Hackathon #AgenticAI

We're hiring for multiple open positions, including these roles on our R&D team: 📙 Senior Backend Software Engineer We're seeking a Senior Backend Engineer who's passionate about building robust data pipelines, integrating with third-party APIs, and working with large-scale data infrastructure. Ideally, you're performance-minded and have experience with modern ETL patterns, workflow orchestration engines, and building API clients. This is a great opportunity to be a member of an exceptional engineering organization with high agency: where you can help shape the features your team builds and how your team builds them. 📕 Senior Backend Engineer - APIs, Data, and Scale We're looking for a senior backend engineer who loves untangling complexity and building systems that scale. This role demands someone who can navigate complicated data models, trace dependencies across systems, and design solutions that are both powerful and maintainable. You'll work with data from project management tools, version control systems, and collaboration platforms to build a complete picture of how engineering work flows through an organization. 📘 Senior Platform Engineer We are looking for a Senior Platform Engineer who views "manual effort" as a bug. You aren't here to just watch dashboards - you’re here to build the pavement as the jet is taking off. We’re scaling fast, dealing with massive data sets, and we need someone who is obsessed with performance, resilience, and the art of reducing toil. See all open jobs here and apply today: https://lnkd.in/gXiHKR2m

Hackathon, May 21 in Boston: build your own workflow or visual editor for a malleable AI developer workspace, in 90 minutes Hosted by Nimbalyst and Jellyfish at Pillar VC as a warm up to BosTech Week. Bring a workflow that annoys you or a file type that deserves a real editor. By the end of the night, you'll have a working extension you can keep using in an open-source visual workspace for building with Codex, Claude Code, Opencode. A few things already shipped as extensions: - Astro website editor - Visual git log - Mindmap - Slides editor - 3D object editor - Mockup builder, diagram canvas, data modeler, spreadsheets If you've ever thought "this file type deserves a real editor" or "this workflow shouldn't live in five tools," come and build a solution for yourself and team. 📍 Pillar VC, Boston 🗓️ Thursday, May 21 🍕 Pizza and drinks on us ⏩️ Limited spots: https://luma.com/xxf112kf

Blitzy CTO Sid Pardeshi joins Jellyfish Advisor Adam Ferrari on May 19 for an exclusive deep dive into the 2026 State of Engineering Management Report that drops this Thursday. Sneak peek into this year’s report's findings: - 64% of teams are already seeing 25%+ productivity gains from AI - Top adopters are achieving 100–150% improvements in output - And yet, most organizations still struggle to measure impact and prove ROI Register to learn how your organization can learn from and apply those insights: https://lnkd.in/g3kU46db

Arcadia's VP of Engineering, Jim Turpin, shares that "Jellyfish really helps me understand where my teams are working and how effective our rollout of AI has been. It helps me understand our spend from what can and can’t be capitalized, and where we’re spending our time, from KTLO and infra to features. It has been great for us to understand how our AI tooling rollout has gone and provides clear visibility into team work." Start leading AI transformation with clarity: https://lnkd.in/gTGVKAym

We asked folks the question "how has AI changed the structure of your team?" Dawshiek Yogathasar (OPTIMAZE), Vincent Tsugranes (Red Hat), Meryem Arik (Doubleword), Matthew Persons (Oumi), and Dorian Matić (Ingemark) share their takes in this video. And it's making us curious to know...how has your team structure changed?