1. agent mission statement

context: a system exists where non-technical people build and deploy autonomous agents. these users create agents to perform complex, multi-step tasks in digital environments.

problem: users can articulate a desired outcome but cannot predict the exact sequence of actions required. the agent needs a durable, orienting principle to guide its independent actions.

solution: provide a single, natural language text field for the user to write the agent's mission. this statement defines a high-level goal and its purpose. the system translates this mission into a primary objective function. during operation, the agent continually refers to this mission to generate, evaluate, and prioritize potential steps.

living pql
context: starting a project in an unknown domain.
problem: design artifacts describe components, missing the life they enable.
solution: an ai analyzes ethnographic data. it synthesizes a living language of human patterns and qualities. this grammar guides all creation.

agent-assisted ethnography
context: conducting field research.
problem: a researcher's perception is limited and biased.
solution: a researcher pairs with an ai agent. the agent records objective data streams. this frees the human to focus on empathy and nuance.

synthetic rituals
context: testing concepts before building.
problem: prototypes test isolated features without integration into life.
solution: generate ai simulations of user routines from the living pql. builders use rough tools inside these simulations to feel how they fit existing life patterns.

adapting shells
context: designing a system's interface.
problem: static layouts impose a single workflow on everyone.
solution: an embedded ai agent observes personal use. it adapts the tool's layout to fit an individual's habits, creating a responsive shell.

the pattern loom
context: making sense of live user feedback.
problem: feedback is fragmented and lacks actionable context.
solution: an ai weaves all feedback into a whole. it maps problems to the living pql, showing where the system breaks a core human pattern.

reflective loop

you are building a system that requires a high degree of accuracy or adherence to a specific format from an llm. a single generative step is insufficient.

the output of a language model is inherently stochastic and may not meet quality standards.

construct a two-part process within a single logical step. a generator llm produces an initial response. a separate critic llm then reviews this response against the original prompt and explicit criteria. the critic's feedback prompts the generator for a refined output.

adaptive dispatch

your application must orchestrate multiple, distinct tools or specialized agents to fulfill a user request. the correct sequence of tools is unknown in advance.

designate a specific llm as a router. this router receives the user's intent and a manifest of available tools, each with a clear description of its capability. the router outputs the name of the next tool to call.

conversational kernel

you have a system that engages in extended, stateful interactions. the history of the interaction is essential for context.

appending the full conversation history to each new prompt is inefficient and costly.

maintain a running, condensed summary of the interaction. after each exchange, use an llm to integrate the new information into this summary. this living document passes to the next turn as the basis of context.