Designing Systems Where Agents Are First-Class Users. The discipline of building tools, APIs, and workflows that agents can operate reliably.
UX → DX → AX. The third user class has arrived.
Mind · Body · Soul
Three systems that give AI agents memory, execution, and identity. Not bigger models—better constrained ones.
Each layer solves a distinct problem. Together they compose into agents that remember, execute, and learn.
Episodic + structural memory. Frames capture what happened. Atlas indexes by policy-aware neighborhoods. lexmap.policy.json defines module boundaries.
✓ Open source · npm @smartergpt/lex · v2.0.0
Explore Lex →Deterministic orchestration. Reads Lex contracts, computes dependency order, runs uniform gates, builds merge pyramids. Predictable outcomes.
Private · Built on Lex 2.0.0 contracts
Learn about LexRunner →Behavioral identity. Turns repeated corrections into scoped rules via Bayesian reinforcement. Persona survives sessions and model upgrades.
Research · 87% precision · <500 token overhead
Read the paper →The Control Deck orchestrates work through a structured planning cycle. The model never directly edits—it plans and allocates.
| Phase | What it does | Math |
|---|---|---|
| Perceive | Parse intent, extract features, classify risk | $\vec{I} = g(h(x), s(x))$ |
| Integrate | Pull Lex state, check version contracts, validate scope | $\phi(\vec{I}, \mathcal{L}, \mathcal{S}, \mathcal{C})$ |
| Layout | Design task graph, assign providers, set budgets | $\lambda(\mathcal{I}, \mathcal{P}, \mathcal{B})$ |
| Orchestrate | Emit plan to LexRunner, configure gates | $\omega(\mathcal{T})$ (pure) |
| Track | Consume Receipts, decide: accept / re-plan / escalate | $\tau(\mathcal{O}, \mathcal{R}, \mathcal{C})$ |
📖 Full documentation in the Lex repo (research papers use "Control Stack")
Formalizes Atlas Frames: fold-radius neighborhoods for adjacency-bounded recall. Instead of dumping entire codebases into context, retrieve only the policy-aware neighborhood around touched modules.
The mathematical foundation for Lex's memory model.
Persistent AI agent identity through frequency-weighted behavioral memory. Bayesian confidence model turns repeated corrections into scoped rules. Hierarchical scope precedence. Bounded prompt injection.
Evaluation results: 87% precision, 80% recall at cosine 0.85 threshold. Prompt overhead <500 tokens for 10-20 active rules.
A rigorous, open scientific study comparing how different AI models (Claude, GPT, Gemini, open-weight) comprehend and operationalize the Agent eXperience framework. Identical prompts, identical control documents, structured responses.
Research questions: Comprehension fidelity, actionability, self-identification accuracy, cross-model variance.
Designing Systems Where Agents Are First-Class Users — The discipline of building tools, APIs, and workflows that agents can operate reliably. Five principles, a maturity model, and why AX matters.
A Frontier Model's View of the Control Stack — A fictional open letter describing the guardrails and control stack that models actually want. Covers scope, gates, modes, and receipts.
Six interlocking concepts: Receipts, Gates, Modes, Policy Surface, Epistemic Guardrails, and Scope & Blast Radius.
How Lex went from alpha to stable contracts. The real story: not features added, but boundaries drawn. What makes 1.0.0 trustworthy.
Real-world case study analyzing agent behavioral failures and how LexSona addresses them through frequency-weighted learning and scoped rules.
Comprehensive review of related work: MemGPT, ChatGPT Memory, RLHF personalization, and academic approaches to behavioral memory.
Models don't need superhuman judgment—they need guardrails that encode the judgment their principals already have about risk, scope, and acceptable behavior.
Give them explicit scope, required gates, structured modes, and audit trails. You get work that's reviewable in minutes, reversible with git revert, and trustworthy because the receipts are there.