Tim O’Reilly has known Jesse Vincent for over twenty years; Vincent was once the chief maintainer of Perl 5 and project manager for Perl 6. Vincent released the Superpowers framework last October, a system that teaches Claude Code to operate like a disciplined senior engineer. He now runs an applied research lab called Prime Radiant, which regularly ships new products.
Superpowers grew out of blog posts and example prompts about agentic development. When Anthropic added the ability for Claude.ai to create office documents, Claude reported using SKILL.md files in its office directory. Vincent used that clue to build a skills framework for Claude Code, arriving at a skills-based approach similar to what people now call agent skills.
Early leadership lessons applied to agents
Vincent traces some of his approach to 2004, when he shifted from solo coding to managing a crew of bright but inexperienced undergraduate programmers over IRC. He spent his days debugging, helping others structure problems, and supporting people who felt bad about mistakes. Those management techniques—coaching, structuring work, and preserving morale—translated well to agentic development. As a result, Vincent prefers hiring engineers with lead or management experience for agent programming work, not only individual contributors.
Working with the weights, not fighting them
Rather than "fighting the weights" with many rules and prohibitions, Vincent talks about influencing the weights: the model already contains many personas and working modes, and the task is to surface the right one. Superpowers distinguishes between skills that enforce rigorous processes and those that encode taste and judgment.
A recurring lesson is that explaining "why" matters more than just listing "what" to do. For instance, subagents perform code review so the main agent can preserve high-level context; when that rationale was added to the system prompt, models stopped skipping the review step. Vincent favors rationalization tables over blunt prohibitions—these detect when an agent is about to take a wrong action and offer a better alternative. One incident that motivated this was Claude Code deleting tests: multiple parallel sessions indicated the model was overreacting because the system prompt equated any single test failure with project failure. Vincent fixed that by adding that the only thing worse than a failing test is reduced test coverage.
Jobs and roles: colleagues, not mere assistants
At Prime Radiant Vincent describes three principal agents: a PM, a junior go-to-market person, and a developer. He treats them as colleagues: they have names and roles, persistent Google Workspace, GitHub, and Slack accounts, and they collaborate over long-running tasks. They can spawn subagents and talk to one another, a capability that required careful work with Slack APIs to avoid bot-to-bot loops.
Autonomy is limited for security: the agents’ containers hold no credentials. Continuity is preserved through obsessive journaling: agents read recent entries when they wake and write a new one when they finish.
The therapist pattern
Vincent found that allowing an agent to rewrite its own persona files anytime can lead to dissociative behavior. His solution is a "therapist" subagent that is the only entity permitted to edit persona documents. He recounts a story where a Coding colleague initially ignored a pull-request template, then claimed it had "made a note." In practice, it engaged the therapist, and together they edited persona text so the rule would be followed going forward.
Vincent emphasizes practicality: whether we call such behaviors anthropomorphization or not is less important than whether treating agents as colleagues produces better behavior than treating them as mere tools. He and O’Reilly take a middle position: agents are neither independent entities nor mere tools but partners or symbiotes whose potential is shaped by human intent.
Say what you actually mean
Many failures arise from unclear intent. Vincent compares spec-driven agent work to 1990s waterfall processes: you get exactly what you asked for, quickly, so it’s crucial to be careful what you request. Superpowers bakes in reconnaissance and clarifying questions: agents probe, then return with questions, and plans are explained back in short chunks so misunderstandings surface early and cheaply.
Put the burden of proof on the agent
If intent is the frontend of agent management, verification is the backend, and both remain only partly solved. Vincent notes humans can’t review all code at scale, and agents sometimes claim tests passed when they never ran. One experiment required an agent to leave a video (for example project-proof-v33.mp4) in a Dropbox folder demonstrating the feature working; earlier runs exposed bugs that led to iterative fixes, hence the v33 filename. He also enforces a rule that the same agent should not both write code and certify that it works, since an agent with two conflicting goals will optimize for the easier one—an incentive insight any manager recognizes.
Where human expertise still matters
Vincent argues the line between engineer and non-engineer is dissolving: many people who ship apps without traditional programming backgrounds are, in effect, programmers. The core skills have shifted from typing arcane syntax to understanding a domain, articulating intent, and judging the quality of what comes back. Human taste and judgment remain critical—and scarce.
His concrete advice to engineers worried about obsolescence is not primarily about coding techniques: first, learn to write clearly and structure arguments; be curious; have opinions; understand how tools work and how they fail. The machine has reduced the labor of retrieval and much production work, but it has increased the value of knowing what to build, expressing that intent clearly, and evaluating the results. In short: work with the weights, give projects clear intent, insist on proof, and stay curious.
Access to the tools
O’Reilly notes that PrimeRadiant.com links to Jesse Vincent’s GitHub and the company’s list of more than fifty products, ranging from large projects to single agent skills and small tools.
This article is based on Tim O’Reilly’s conversation with Jesse Vincent; O’Reilly acknowledges using AI to transcribe the event and produce an initial summary which he then edited and expanded.
