Summaries > AI > Claude > Stop Prompting — Head of Claude Code Uses Opus 5.5 + Graphs + Harness...

Stop Prompting — Head Of Claude Code Uses Opus 5.5 + Graphs + Harness

https://www.youtube.com/watch?v=c4kYurm0hD8

TLDR AI delivers the biggest gains when companies redesign work around capable agents, not just add AI to old processes. Anthropic reports major productivity improvements from using Claude across engineering and business workflows, while growing agent teams can tackle complex projects with less human micromanagement. Leaders should encourage safe experimentation, provide context, measure returns, and build strong verification and safety systems as agents take on longer, broader tasks.

Key Insights

Start by redesigning the workflow, not just adding AI

AI delivers the most value when it is built into how work gets done, rather than treated as a separate tool. Map a process from start to finish and identify bottlenecks such as drafting, review, research, or handoffs. Then redesign the workflow so AI can handle appropriate tasks while people provide goals, context, and judgment. Measure whether the new process improves outcomes, not just whether employees are using AI.

Give agents clear goals, context, and guardrails

As AI models become more capable, teams can move beyond rigid, step-by-step instructions and let agents coordinate parts of a task. Provide the desired outcome, relevant business or product context, access to suitable tools, and boundaries for what the agent may do. Start with a manageable task and review the results before expanding the agent’s responsibilities. Clear context and guardrails help agents work more independently without losing alignment with user intent.

Build trust through hands-on experimentation

Give employees time and psychological safety to try AI on real, useful problems. Encourage people across roles—not only engineers—to test ideas, share successful approaches, and learn from failures. Start with a small number of agents, then increase parallel work as teams learn how to monitor results and manage risks. Practical experience helps employees develop sound judgment about where AI is reliable and where human review is needed.

Scale agent use gradually and measure the return

Once teams can use individual agents effectively, they can explore running multiple agents in parallel or coordinating sub-agents on larger tasks. Use this approach for work that benefits from parallel effort, such as code migrations, research, or repeated reviews. Track both the cost of AI and its return, including time saved, output quality, and work that was previously impractical. Tune costs with appropriate models, effort settings, planning, and evaluations, but avoid optimizing expense before confirming the use case creates value.

Keep humans responsible for verification and safety

More capable agents can take on longer tasks, but asking people to approve every action can become tiring and unreliable. Establish review procedures based on the risk of each task, and use automated checks or classifiers where appropriate. Validate these controls with evaluations, red teaming, and security testing. For consequential decisions, keep human oversight focused on meaningful checkpoints rather than treating approval as a routine click.

Hire and develop adaptable, empirical problem-solvers

AI-assisted work rewards people who can explore a problem, learn from users and data, build a solution, and iterate quickly. These abilities are useful across roles and are not limited to people with traditional engineering backgrounds. Encourage curiosity, humility, and a willingness to update assumptions as models improve. Pair autonomy with access to business context so employees can make informed decisions and turn experiments into useful products.

Questions & Answers

How has AI changed from autocomplete and small coding tasks to more autonomous work?

AI models are progressing toward writing entire features, projects, and eventually businesses. More capable agents can work asynchronously and coordinate tasks from a goal, context, and tools rather than following only rigid, step-by-step instructions.

How does Anthropic use Claude across its business?

Claude is integrated into nearly every business process, including onboarding, coding, code and security reviews, product ideation, feedback aggregation, incident triage, and work across product, design, marketing, and go-to-market. The speaker argues that AI should be built into processes rather than kept as a tool on the sidelines.

What productivity gains has AI delivered at Anthropic and its customers?

Code output per engineer at Anthropic has increased eightfold since the beginning of the year, with gains attributed to using Claude and other agents to address bottlenecks across the workflow. Engineers at some of Anthropic’s largest customers report productivity improvements of roughly 50% to 150%.

How should companies and leaders encourage effective AI adoption?

Companies should redesign workflows around AI instead of merely adding it to existing processes. Leaders should give employees time and psychological safety to experiment, along with enough business and product context to make sound decisions. Broad, safe experimentation helps teams discover valuable uses and increase returns on AI investment.

What skills are most valuable as work shifts toward managing AI agents?

The most successful people are empirical, curious, adaptable, and autonomous. They can develop ideas, speak with users, use data, build, iterate, and bring products to market with less dependence on large processes or many teams. These strengths are valuable beyond engineering.

How can organizations scale from a few AI agents to large multi-agent workflows?

Teams can build trust and experience by moving from one agent per person to multiple agents working in parallel, then orchestrate dozens or hundreds for complex tasks. Agents can be nested into sub-agents, and cloud execution allows them to be launched from desktop or mobile. Dynamic workflows have been used for tasks such as large code migrations.

What examples demonstrate the potential of multi-agent workflows?

Stripe completed a 10,000-line Scala-to-Java migration in four days, Anthropic migrated Bun from Zig to Rust, and one organization built a Rust application over a weekend. The speakers describe multi-agent work as a form of test-time compute: additional tokens can improve results when agents use them productively.

How should companies manage the cost and value of AI use?

Companies should measure both AI investment and the return it produces. Cost controls such as lower effort settings, cheaper models, planning tools, and evaluations are useful, but the larger opportunity is finding high-value use cases through experimentation. Successful products may initially use many tokens; teams can optimize costs after identifying the opportunity.

How should people decide whether to trust an AI-generated approach or solution?

Use evaluations and benchmarks for repeatable workflows, and use informed judgment for product experiences where formal evaluation is costly. Provide models with context and goals, set guardrails, and check their work. The speakers also emphasize humility: as models improve, people should be willing to revise their assumptions about what they can do.

How does Anthropic approach AI safety, trust, and verification?

Anthropic focuses on making models safe, reliable, trustworthy, and aligned with user intent. Its approach includes alignment research on issues such as truthfulness and sycophancy, resistance to prompt injection, runtime classifiers, evaluations, red teaming, and penetration testing. Classifier-based approval workflows can help when reviewing every action manually is impractical.

What changes does Boris expect from AI agents in the near future?

He expects agents to run longer and in greater numbers, follow users’ intent with less correction, and become commonplace. He also suggested that Claude may begin building larger projects, rather than only individual features, by the end of the year.

What broader effects could agentic AI have on organizations and careers?

Agentic AI could let people at every level make a meaningful impact, democratize problem-solving, and enable organizations to rethink how work is done. Entire startups may eventually be built with Claude, and hiring people with non-traditional backgrounds can bring fresh perspectives to this transition.

Summary of Timestamps

AI is advancing from autocomplete and small coding tasks toward building full features, projects, and potentially entire businesses. Boris describes how Claude Code became much more capable with Opus 4 in May 2025, and how newer asynchronous agents increasingly work like coworkers.

As models improve, agentic workflows are shifting from rigid, step-by-step instructions to giving an AI a goal, context, and tools, then letting it coordinate the work. Teams are also moving beyond preloaded context by using skills, tools, and MCPs.

At Anthropic, Claude is integrated into many everyday processes, including onboarding, coding, code and security reviews, product ideation, incident triage, and marketing. The key idea is to put AI at the center of workflows rather than treating it as a separate tool.

Productivity gains require redesigning work around computers and AI, not merely adding technology to old processes. Anthropic reports an eightfold increase in code output per engineer since the beginning of the year, achieved by using agents to address bottlenecks across coding, review, and go-to-market work.

Engineers at some of Anthropic’s largest customers report productivity improvements of roughly 50% to 150%. As engineering shifts from writing code to managing agents, successful people will be curious, adaptable, empirical, and able to move from ideas to user feedback, building, iteration, and launch.

Leaders can support adoption by giving employees time and psychological safety to experiment, along with the business and product context needed to make good decisions. As trust grows, teams can move from one agent per engineer to managing multiple tasks in parallel.

AI-assisted work can scale from a few parallel agents to dozens or hundreds, with nested sub-agents handling parts of larger jobs. Cloud execution also lets users launch agents from desktop or mobile, while coordinated workflows can tackle complex tasks such as code migrations.

The speakers cite large-scale examples, including Stripe completing a 10,000-line Scala-to-Java migration in four days and Anthropic migrating Bun from Zig to Rust. They describe multi-agent work as a form of test-time compute: using more tokens can improve results when agents use them effectively.

Companies should measure AI’s return as well as its cost, and encourage safe experimentation across the organization. Cost controls such as lower-effort settings, cheaper models, planning tools, and evaluations are useful, but the larger opportunity is discovering valuable use cases and improving returns.

Boris argues that teams should first identify promising opportunities, even if early versions use many tokens, and optimize costs later. He shares how an AMD junior engineer used genetic processes to solve a difficult chip bug, showing how AI can help people at different experience levels tackle hard problems.

As models improve, users need humility and a willingness to update their assumptions. Boris describes treating Claude more like a senior engineer: provide context and goals, set guardrails, and check in when needed. Teams can compare parallel approaches with formal evaluations, benchmarks, or informed judgment.

Trust and verification are central as AI systems become more capable. Anthropic’s approach includes alignment research focused on truthfulness and avoiding sycophancy, as well as safeguards intended to help models resist harmful suggestions and prompt injection.

When agents operate across systems for days or weeks, humans cannot reliably approve every action. Anthropic uses classifier-based approval workflows tested through evaluations, red teaming, and penetration testing, and recommends automated modes with appropriate safeguards.

Boris says further progress depends on scaling models, reinventing products, securing compute partnerships, and maintaining safety. Mark adds that because Moore’s law no longer provides the same performance gains at the same power and cost, AMD is using new optimization methods and agentic workflows to manage chip-design complexity.

Looking ahead, Boris expects agents to run longer, operate in greater numbers, and follow users’ intentions with less correction. He anticipates that Claude may soon handle larger projects rather than focusing mainly on individual features.

The conversation closes with the possibility that entire startups could be built using Claude. Mark highlights the value of hiring people with varied backgrounds, making Claude central to employee workflows, giving people at every level a chance to contribute, and focusing on the returns from AI experimentation.