Chapter zero / The interface of work
People learned which sites to search, which data to trust, which buttons to press, and how to assemble the result. Software waited for the next instruction.
Now the person describes an outcome. An agent can gather context, choose tools, execute steps, and return at the moments that need judgment.
Applications expose functions. The person carries the goal, context, sequence, and quality bar from screen to screen.
Applications become capabilities. The agent carries state across them while the person supplies purpose, expertise, and approval.
Web 4.0 does not remove professional work. It moves effort away from serial application operation and toward designing, teaching, governing, and improving the system.
Improvement usually means the person becomes faster at research, navigation, templates, and production.
The durable advantage becomes the quality of the workflow, context, skills, permissions, and expert feedback—not prompt cleverness alone.
We are moving from people operating software to people designing how software pursues outcomes.
01 / The Web 2.0 operating system
Each application owned a slice of the work. The human remembered the goal, moved information between systems, and decided what happened next.
Click any company or platform to see what it contributes. The stack is powerful, but the person translates between layers.
02 / The Web 4.0 operating system
The human defines the goal and the boundaries. An agent runtime assembles context, chooses tools, executes work, and pauses when judgment or authority is required.
Click any company, product, or protocol to see its role. The new stack carries state while human authority stays explicit.
The core story / Four parts
First, see where the work and judgment live today. Then rebuild the same job around goals, context, skills, tools, and a controlled agent runtime.
Every source, decision, and handoff passes through the person doing the work.
Human time is consumed by operating the workflow.
Human time concentrates on differentiated expertise.
The system underneath the shift
The new stack is the work of turning professional practice into workflows, skills, context, governed tools, evaluations, and feedback loops that improve with use.
Corrections improve the skill, context, and evaluation set for the next run.
The mental model: The model supplies capability. The team builds the workflow, skills, context, permissions, evaluations, and feedback loop that make capability useful.
Value moves from answering questions to owning the path between intent and completed work.
04 / The work stack
This is the work that expands as execution becomes agentic. Click each surface to see what teams will design, refine, govern, and improve over time.
Domain expertise, examples of excellent work, judgment, trade-offs, corrections, and authority.
OWNED BY THE PROFESSIONALWorkflow state, modules, context retrieval, tools, permissions, checkpoints, and output generation.
HELD BY THE HARNESSSkill versions, curated context, better tools, stronger evaluations, policy, and—only when justified—model adaptation.
DRIVEN BY EXPERT FEEDBACKRead it as an improvement loop
Past work informs context. Context and expert practice become skills. Skills run inside workflows. Human corrections improve the next version.
06 / Observable execution
This executes the one-pager system from Build: skill selection, permissioned sources, agent modules, context assembly, evaluation, human correction, and feedback all light up from top to bottom.
Sources are reconciled and citations are attached. A human challenges the thesis before the briefing is published.
Advance one step at a time. Each stage explains what the agent does, which tools are called, and where the individual becomes involved.
The model did not “run the business.” It operated inside a bounded harness: scoped identity, explicit tools, typed inputs, policy checks, durable checkpoints, and an auditable output.
05 / Builder's guide
A concrete walkthrough: begin with a manually produced investment one-pager, extract the team's practice, create a reusable skill, connect trusted sources, and improve the system through expert review.
Select a control to see the default architecture and the question the team should resolve before expanding autonomy.
What is deterministic, what is model-decided, and where does the run stop?
Which irreversible, sensitive, or low-confidence actions require review?
Does every action use least privilege, short-lived credentials, and per-step policy?
What must be retrieved now, with provenance and freshness attached?
What should persist across turns, runs, or customers—and what must expire?
Which systems are read-only, writable, authoritative, or off-limits?
Can you replay the plan, context, tool inputs, outputs, latency, and cost?
Are retries bounded, writes idempotent, and compensating actions designed?
Can untrusted content alter instructions, leak secrets, or expand authority?
Does task value exceed inference, integration, review, and exception cost?
07 / Companies
A focused map of the companies helping teams orchestrate agents, serve models, connect tools, manage context, evaluate behavior, and govern production systems. Vertical applications are intentionally separated from this infrastructure view.
Illustrative, not exhaustive. Category boundaries overlap; ownership and company status can change.
Investment lens
Durable value is likely to accrue where a company owns one or more hard-to-replace assets—not merely where it wraps a model.
The one-slide takeaway
“The model is the brain.
The stack is the operating system for action.”
Pick work with judgment, messy inputs, and measurable value—not a chatbot looking for a job.
Give the agent enough authority to be useful and enough constraints to be trusted.
Reliability compounds when runs become inspectable data, evaluations, and better policy.
This is an educational synthesis, not a definitive taxonomy or investment recommendation. Company examples are illustrative. Core architecture patterns were cross-checked against primary documentation.