--- title: "Typed-Effect Kernels for Reliable Agent Systems" subtitle: "K-12, G-9, and M-7 · working preprint" author: "Ryan Hunter, with AI research collaborators" date: "August 2025 · public source edition September 2026" abstract: | Language-model "agents" have outgrown ad-hoc prompt chains. Production deployments demand typed contracts, auditability, policy enforcement, and cross-organization interaction. We introduce **K-12**, a **typed-effect episode kernel** that captures the environment and laws for one bounded run; **G-9**, a **graph kernel** for typed coordination (routing, scheduling, durability); and **M-7**, a **market/exchange kernel** for cross-organization capability leases, metering, and settlement. The trio is realized in the **Effect TS** calculus (Context Tags + Layers), enabling formal reasoning and hot-swappable back-ends. We provide normative laws, operational semantics, evaluation rubric against LangGraph-style workflows, and a conformance harness. This paper consolidates the specification, design rationale, and mappings to implementation. --- *This is an archival preprint. The later specifications and formalization work refine parts of it; read the current research essay for the operating argument.* **Version:** 0.1.2 (Spec Freeze) • **Date:** 2025-08-08 **Effect Generic Order:** **A-E-R** (`Effect`) **New in v0.1.2.** We add canonical JSON schemas, end-to-end idempotency/"exactly-once" semantics, explicit OTel mappings, a minimal monitor grammar, structured capability/ABAC, Memory/ Ledger envelopes with erasure/tamper-evidence, and transport bindings for λ streams. ## 1 Introduction Agent frameworks have proliferated, yet most blur three distinct concerns: 1. **Local semantics** of one bounded run (what we call an *Episode*). 2. **Coordination** of many runs across components and teams. 3. **Economics/governance** across organizational boundaries. We separate these into three kernels: * **K-12 (Episode Kernel):** Purity of control (δ/λ), explicit effect slots (Tools, Memory, Quota, Policy, Explain, Locale), and **streaming** outputs. * **G-9 (Graph Kernel):** Typed nodes/edges, Router/Scheduler/Ledger separation, guarded cycles, and metrics/monitor hooks. * **M-7 (Market Kernel):** Directory, attestation, contracting, metering, settlement, dispute, reputation. The architecture mirrors proven systems patterns: micro-kernel per run, orchestration fabric above, and market protocols outside trust boundaries. ## 2 Background & Goals * **Typed Effects.** Effect TS models services as Context Tags and compositions as Layers, cleanly separating *ports* (interfaces) from *adapters* (back-ends). * **LLM Agents.** We treat LLMs as tools inside Episodes: Observe/Orient may use embeddings and search; Decide may invoke a chat model; Act executes tools or sub-workflows. * **Goals.** (i) Soundness (typed IO, replayable traces), (ii) Governance (quota, policy, locale), (iii) Scalability (routing/scheduling), (iv) Inter-org operation (markets). ## 3 K-12: The Episode Kernel (See spec for details.) We define the stream chunk envelope (text/json/tool-call/result/error/system), cancellation/back-pressure, transport bindings (SSE/gRPC), error taxonomy, and right-to-erasure. ## 4 Persona: Sequencing Episodes Persona is a tiny controller that decides when/how to run another Episode or delegate to a sub-graph; it lives above K-12. We formalize Transcript/Budget shapes and decision idempotency. ## 5 G-9: The Graph Kernel Axioms on typing, cycles (guards), and routing idempotence. Options cover conditional edges, QoS & partitioning, lineage, human-in-the-loop, and GraphCall (trusted, in-org). We specify scheduling, parallel/join semantics, adapter nodes, and idempotency interactions across R and L. ## 6 M-7: The Market/Exchange Kernel (Alpha) Directory, attestation, offers/contracts, metering, settlement, dispute, reputation. Laws to follow with early adopters. ## 7 Operational Semantics Episode and Graph steps; idempotency keys propagate into L and R enabling safe retries. Ledger is tamper-evident; Memory uses LWW-element-set with tombstones for redaction. ## 8 Observability & Telemetry IDs on spans/ledger; low-card metrics labels with exemplars keyed by trace/span; OTel mapping table enumerated in the spec. ## 9 Evaluation Rubric Parity with LangGraph-style workflows; reliability under faults; governance checks; erasure correctness; observability correlation. ## 10 Security, Privacy, Locale Least privilege; ABAC capabilities; Λ subset discipline; sandbox & egress control; RAG poisoning mitigations; embedding privacy; rotation of V leases. ## 11 Related Work LangGraph, Airflow/Temporal, ReAct/OODA, Effect/ZIO calculus. ## 12 Limitations & Future Work M-7 laws; formal monitors; TLA+ models; reference adapters. ## 13 Conclusion K-12, G-9, and M-7 provide a principled, composable foundation for reliable agent systems. ## Appendices See `spec/` documents for norms, schemas, and conformance suite. ## References and context - Shunyu Yao et al. 2023. [ReAct: Synergizing Reasoning and Acting in Language Models](https://arxiv.org/abs/2210.03629). *ICLR*. - [LangGraph reference](https://langchain-ai.github.io/langgraph/reference/). LangChain. - [Temporal workflow execution documentation](https://docs.temporal.io/workflow-execution). Temporal Technologies. - [Managing Layers](https://effect.website/docs/v4/requirements-management/layers). Effect documentation. ## Source edition notes This edition preserves the August 2025 working preprint, with the unfilled release-commit field removed. The companion specifications appear in *The Kernel Trilogy: Source Specifications*. It is a research artifact, not a certification of every mathematical claim in the draft.