The Post-Model Era
Models are commoditizing on a schedule; skill libraries, audit trails, and domain knowledge are what now compound into a durable moat.
WiseWare blog
The Procedural Substrate
How AI systems moved from prompts as product to runtime environments, trajectories, and durable operational memory.
Skills, Not Agents
Persistent agent identities are a transitional shape; the durable unit is the skill, composed fresh for each task and dissolved when the work is done.
Why Small Prompts and Rich Environments Win
Procedures encoded in tools, skills, and sandboxes are reliable in ways a 4,000-word system prompt never was — and they survive a model swap.
The Hidden Architecture of Claude Code, Codex, and Pi
Three harnesses with near-identical feature lists embody three different philosophies about where intelligence lives and how much autonomy belongs to the agent.
Coding Agents Are Really Workspace Operating Systems
Filesystems, shells, sandboxes, and test runners determine what an agent can do — the model is a component inside that environment.
The Death of Full Autonomy
Why the agents that won production shipped checkpoints calibrated to consequence instead of chasing unsupervised loops.
Context Engineering Replaced Prompt Engineering
Why retrieval, focus, memory layers, and information shape decide more about agent quality than wording ever did.
Why Agent Evals Became Behavioral
How replay, recovery analysis, and trajectory scoring replaced the spreadsheet of inputs and expected outputs.
Trajectory Is the New Interface
Why showing the work became the work, and what changes when the path matters more than the answer.
The Large Procedural Substrate
Why naming the layered system around the model matters, and what its five layers look like in production.
Why Modern Agents Behave More Like Operating Systems
Scheduling, permissions, isolation, drivers, syscalls, audit logs — the harness keeps rediscovering primitives the systems world settled decades ago.
Progressive Disclosure Is the Missing Primitive
A capability the model cannot see is a capability it cannot misuse, which turns out to be a stronger safety guarantee than any instruction.
From System Prompts to Skills
Once procedure lives beside the model rather than inside it, capability authoring becomes a job for domain experts instead of prompt engineers.
The Rise of the Agent Harness
Six pieces of machinery that every team kept rebuilding badly, until the field admitted the harness was the product and the model was the engine.
An Agent Is Not a Chatbot With Tools
The dividing line is whether the system keeps going when the user walks away, and every architectural choice follows from that one fact.
Agents Changed the Unit of Intelligence
Once the artifact is a trajectory rather than a completion, per-step quality stops mattering and joint quality across the path takes over.
The Shift From Prompt-Centric to Runtime-Centric
The instructions get shorter as the environment gets richer, and that single inversion is what separates 2023 demos from 2025 systems.
Why Prompt Engineering Hit a Wall
Five failure modes that turned every long system prompt into an unmaintainable monolith, and why the fix had to be architectural rather than editorial.
Chains Were Our First Attempt at Structure
LangChain gave the industry a vocabulary for steps and routers, then got mistaken for the building rather than the scaffolding holding it up.
When the Prompt Was the Product
How a 4,000-word text file briefly became the most valuable artifact in AI engineering, and why a global namespace was never going to hold.