Agentic Engineering
I do not think AI replaces engineers. It supercharges them, and in doing so it moves the bottleneck. When code generation gets cheap, more code gets written, not less, and the expensive part becomes reviewing, verifying, and integrating it. That is the through-line in everything I write here: the work does not disappear, it relocates.
Once you accept that, the org chart starts to move. Engineers become managers of agents, the leverage shifts up a level, and the questions that matter are about specs, documentation, and evaluation rather than keystrokes. The teams that win are the ones that measure the right thing instead of optimizing token counts, and that build stability around whatever model is current rather than betting the product on a single one.
And almost all of it is a product engineering problem, not a research one. The hard parts are latency as a feature decision, knowing your user is not you, and resisting the urge to ship another chat box. The posts below are the running log of working through that in production.
Reading path (25 posts)
- Adopt the Agent. Build the Loop. -- Everyone is debating Cursor vs Claude Code. Wrong fight. The agent is a commodity. The only thing worth building is the loop around it.
- The Manager Layer is Next -- AI isn't replacing managers. It's exposing which ones were never load-bearing to begin with.
- Tokenmaxxing Is What Happens When You Measure Wrong -- Mandating AI tool usage is correct. Measuring it is a trap. The right move is to measure outcomes, not inputs.
- Relaunching This Blog for the AI Era -- Jeff Adler's engineering blog, eight years of iOS architecture and leadership. Now relaunching for the AI era.
- The Year the Agent Stopped Being a Demo -- 2025 was the year agents went from demo to operational. Some 2024 predictions held. Some didn't. Three bets for 2026, and the gap that hasn't closed.
- AI Is Redefining What \"Senior Engineer\" Means -- Senior engineer used to mean better code. AI is redefining it to mean system design, taste, and judgment. A Dropbox director's take.
- What Two Years of Agentic Coding Did to My Org Chart -- Sixteen months after the agent-managers thesis, here's what actually happened to my org chart. Some held up. Some didn't. Hiring shifted hardest.
- Your Agent's Bottleneck Is Your Documentation -- GPT-5 landed muted. The model layer is stabilizing. The differentiator moved one layer down to your information architecture for agents.
- Voice Is Going to Fail for the Same Reason Chat Did -- Voice as a default surface is failing the way chat did. Voice as an embedded wedge is the actual win, and the early-2025 data already shows it.
- AI Makes You 3x Faster, Then Review Explodes -- AI tools help Dropbox engineers ship 3-5x faster, but without the right processes you just create a traffic jam at code review.
- Reasoning Is a Commodity Now -- DeepSeek R1 and Claude 3.7 made reasoning a commodity in five weeks. The architectural question moves from which provider to where you absorb the cost.
- Should We Write Code for LLMs Now? -- If LLMs increasingly read and modify our code at Dropbox, should our conventions optimize for them instead of humans?
- Your User Is Not You -- The biggest AI product mistake of 2024 was shipping for ourselves. Users don't know what to prompt and bounce on the first reply. Empathy gap is PMF gap.
- Agentic Workflows Are Harder Than You Think -- The gap between an impressive agent demo and production at Dropbox is enormous. Here's what bridging it actually takes.
- Commit Your Specs (We're Almost There) -- o1 is the first model where the spec might be the load-bearing artifact, not the code. I am hesitant. The QA bar this requires is one the industry hasn't built.
- From Copilot to Pipeline -- Cursor and Copilot are the trust-but-verify phase of AI coding. The next phase is a pipeline of agents that writes the code while you review.
- Engineers Are Becoming Agent Managers -- AI is flattening the engineering output curve at Dropbox. Engineers are managing agents now. Org structures need to change.
- Stop Picking Models. Pick What's Stable Around Them. -- Claude 3 dethroned GPT-4. Devin redefined the agent. Teams that hardcoded a model are scrambling. The architecture that survives picks anything but the model.
- Stop Building Chat. Your Users Don't Want Another Box. -- Every B2B company is racing to bolt an "Ask AI" chat onto their product. Most users bounce on the first reply. The chat box is product engineering laziness.
- What I Got Wrong About RAG -- RAG isn't what makes your AI product good -- retrieval quality is. Most teams learn that the hard way by optimizing the wrong layer first.
- Shipping AI When Nothing Works Yet -- The messy reality of shipping LLM-powered features at Dropbox with real users, real latency budgets, and a cost model that doesn't work.
- The Evaluation Gap Is the AI Gap -- Teams that build real eval infrastructure first ship better AI products -- not because they have better models, but because they know when their model is wrong.
- Prompt Engineering Is Not a Job -- Hiring a prompt engineer is a category error. The real job is eval design, statistical rigor, and infrastructure -- and no one is posting that JD.
- Latency Is a Feature Decision, Not an Infrastructure Problem -- How you handle latency determines whether users trust your AI product. Make that decision at design time, not in production.
- ChatGPT Just Made AI a Product Engineering Problem -- ChatGPT launched eight days ago. Mobile engineers thinking "this doesn't affect me" are wrong. The product engineering reset starts here.
Frequently Asked Questions
Does AI replace software engineers?
No. In Jeff Adler's view AI supercharges engineers and shifts the bottleneck from writing code to reviewing and verifying it, so more code gets written, not fewer engineers needed.
What is agentic engineering?
Building software where AI agents do much of the code generation under human direction, which changes the engineer's job toward specification, review, evaluation, and orchestration.
More topics: Engineering Leadership | iOS Architecture at Scale | All posts | About Jeff Adler