Agentic Workers

An agent can chain tools and sometimes recover from earlier failures. Andrew Ng's planning example also keeps the important qualifier: these workflows can be finicky. From his 2024 Sequoia talk. #AgenticAI #AIWorkflows

People expect an assistant that remembers what they said, not just text generation. Andrej Karpathy on the gap between the underlying model and a useful interface. No Priors, March 2026. #AI #ProductDesign

Before increasing autonomy, ask how costly an error would be and how hard it would be to discover. Barry Zhang on read-only access and human review. April 2025 talk. #AI #RiskManagement

Study the foundations, understand the tools, and practice using them in your actual work. Demis Hassabis's advice to students in a June 2025 WIRED interview. #AI #CareerSkills

Shadow mode. Human recommendations. Limited actions with kill switches. Accenture's autonomy ladder makes each step depend on evidence and trust, not a completed project plan. #AIGovernance #AgenticAI

Two accounts can have different email addresses. Karpathy's payment example shows why a working-looking implementation still needs human review of identity and data design. #CodeReview #AIEngineering

Uber's MCP lesson: custom integrations built in silos aren't enough. A dependable production system needs a shared development and deployment lifecycle. #MCP #AIInfrastructure

One practical review loop: give the model its own code back and ask it to check correctness, efficiency and construction. Andrew Ng keeps the caveat: improvement isn't guaranteed. From his 2024 talk. #AIWorkflows #CodeReview

Think in larger units of work: one non-conflicting feature per agent, then review the result. Andrej Karpathy on building a delegation habit. No Priors, March 2026. #AIAgents #SoftwareDevelopment

Not every task needs an agent. Barry Zhang's practical test: if the decision tree is clear, build it explicitly. From his April 2025 talk on effective agents. #AIAgents #Workflows

Build, evaluate, iterate. Accenture argues that AI delivery teams need to turn evidence into stakeholder confidence and report what they learned, not only what they delivered. #AIEvaluation #EnterpriseAI

Being able to automate a task does not settle whether people want it automated. Demis Hassabis uses care and empathy to explain the distinction. His perspective in a June 2025 WIRED interview. #FutureOfWork #AI

What would an agentic-engineering interview actually test? Karpathy proposes a substantial project, real tooling and adversarial attempts to break the result, not just puzzles. #Hiring #AgenticEngineering

Business cases ask useful questions. But can you really know the solution, value and delivery cost before experimenting? Jess Grogan-Avignon on learning the business case through the work. #EnterpriseAI #Experimentation

A first draft isn't the workflow. Andrew Ng explains why outlining, researching and revising beats asking for one perfect answer. From his 2024 Sequoia talk. #AIWorkflows #AgenticAI

Vibe coding raises the floor. Agentic engineering keeps the professional quality bar intact. Karpathy on moving faster without handing away responsibility for the software. #AgenticEngineering #SoftwareQuality