From vision
to execution
and assurance.
AI agent harness means the vendor-neutral control and execution environment around models and agents: context, tools, permissions, state, decision gates, evidence and effects. It does not refer to Harness.io. The layer labels below are navigation anchors for this reference architecture, not an external standard.
Start here: architecture map
Read top-down when you are new to the series. Each layer answers one governance question and links to the best current entry point.
What is the governed system?
Define purpose, responsibility, data, authority, execution and evidence boundaries before choosing products.
Read article ↗L1 · Strategy & PortfolioWhat should become an organizational commitment?
Turn candidate AI use cases into funded commitments with explicit value, risk, ownership and evidence expectations.
Read article ↗L2 · Governed Continual HarnessHow may an agent capability change safely?
Bind experiments to evidence, risk review, independent qualification, accountable approval and controlled promotion.
Read article ↗L3 · Runtime & OrchestrationHow do agents plan, act and progress under control?
Treat context, planning, tools, state and progression authority as part of the effective system.
Read article ↗L4 · Governed Context & KnowledgeWhich information is eligible to become operative context?
Admit evidence and memory by provenance, freshness, classification, applicability and revocable purpose.
Read article ↗L5 · Capability & IntegrationWhich skills, tools and integrations are admitted?
Move reusable capability from ad hoc prompts into versioned, governed contracts with explicit authority and lifecycle.
Read article ↗L6 · Security, Identity & GovernanceWhich boundaries constrain the system?
Apply identity, permissions, guardrails and governance at the layers where risk and consequential effects are created.
Read article ↗Layer-by-layer foundations
Existing layer-aligned articles are ordered by their primary place in the architecture. New Tuesday/Thursday publications will fill the missing layers in sequence.
Capabilities Are Not Permissions
Separate tool discovery, possession, eligibility, invocation and authority, then verify the actual effect.
Read article ↗L4 · Applied controls · 29 Oct 2026Memory Admission: What an Agent Is Allowed to Remember
Promote context into durable memory only with provenance, purpose, authorization, expiry and revocation controls.
Read article ↗L4 · Architecture · 27 Oct 2026Context Is Governed State, Not Just Tokens
Admit context by provenance, purpose, freshness, classification and revocable trust rather than window size.
Read article ↗L3 · Applied control · 27 Oct 2026 (22 Oct slot)Planning, Delegation and Synthesis Without Losing Decision Authority
Bound delegation, preserve conflicting evidence and separate decision authority from coordination.
Read article ↗L3 · Architecture · 20 Oct 2026Central Authority, Distributed Cognition
Separate distributed specialist reasoning from consequential decision authority.
Read article ↗L2 · Applied control · 15 Oct 2026Promotion Is a Control: From Experiment to Qualified Capability
An executable promotion gate binds identity, evidence, authorization, observed deployment and revocation.
Read article ↗L2 · Governed Continual Harness · 13 Oct 2026The Governed Continual Harness: How Agent Systems Are Allowed to Change
Separate experimental improvement from evidence-based qualification and authorized promotion.
Read article ↗L1 · Applied control · 8 Oct 2026The AI/Agent Factory: Promotion, Ownership and Portfolio Gates
Separate portfolio progression from capability promotion and deployment authorization using explicit value, risk, ownership and evidence gates.
Read article ↗L1 · Strategy & Portfolio · 6 Oct 2026From AI Use Cases to a Governed Portfolio
Use cases become portfolio commitments only after value, ownership and evidence expectations are explicit.
Read article ↗L0 · Applied boundary · 1 Oct 2026Architecture Before Tools: Defining the Boundary of a Governed AI System
Draw the governed boundary around purpose, responsibility, data, authority, execution and evidence—not around a product.
Read article ↗L0 · Vision · 29 Sep 2026The Enterprise AI Operating System: What the AI Agent Harness Actually Governs
Define the governed system before selecting frameworks, models or tools: business intent, decision rights, lifecycle, evidence and explicit boundaries first.
Read article ↗L0 · Vision · 1 Sep 2026Toward a Reference Taxonomy for Governed AI Systems
A vendor-neutral map connecting capabilities, architecture, decisions, evidence, risks, controls, responsibilities and lifecycle.
Read article ↗L0 · Architecture boundary · 3 Sep 2026Frameworks Are Not Architecture
Separate durable capabilities, trust boundaries and governance contracts from replaceable frameworks, protocols, models and vendors.
Read article ↗L1/L2 · Operating model · 21 Aug 2026The Agentic GRC Operating Model
Use specialised agents to collect evidence and route decisions while human owners retain authority.
Read article ↗L2/L6 · Policy to control · 21 Aug 2026From Policy Hierarchies to Controls
Design a machine-readable control system that produces usable evidence instead of a document graveyard.
Read article ↗L2/L6 · Governed enforcement · 21 Aug 2026From Paper Policies to Agent-Driven Controls
Turn machine-readable requirements into governed gates, evidence contracts and accountable escalation.
Read article ↗L3 · Runtime & orchestration · 22 Sep 2026The AI Agent Harness Is Part of the Model
Why context, planning, action space and governance determine the effective capability, cost and risk of an AI agent.
Read article ↗L5 · Capability & integration · 16 Sep 2026Skills Replace Prompts
Move recurring AI work from improvised instructions to versioned capability contracts with explicit tools, authority, evidence and lifecycle.
Read article ↗L6 · Security, identity & governance · 8 Sep 2026AI Guardrails Are More Than an Output Filter
A layered control architecture across governance, identity, retrieval, generation, tools, outputs and operations.
Read article ↗Enterprise Assurance · cross-cutting · 8 Sep 2026AI Can Draft the Assessment. It Cannot Own the Assurance Decision.
A practical architecture for AI-assisted GRC: evidence first, governed analysis, independent challenge and explicit assurance authority.
Read article ↗Architecture releases
Release notes describe changes across the whole AI agent harness architecture and are not part of the layer-by-layer article sequence.
AI Agent Harness Architecture v1.4.0: Contract Cognition to Effect
Contracted model-harness units, authority-preserving coordination, governed context, spec-driven delivery, executable verification and pre-effect control.
Read article ↗Architecture release · 1 Oct 2026AI Agent Harness Architecture v1.3.0: Govern the Effective System
Governed state transitions, effective-system capability, inherited policy, skill evolution, external-server trust, cross-layer evidence and outcome-based assurance.
Read article ↗Architecture release · 25 Sep 2026AI Agent Harness Architecture v1.2.0: Bind Decisions to Effects
Approval continuity, transitive authority, governed harness evolution, learned team organization and control-retaining distillation.
Read article ↗Architecture release · 18 Sep 2026AI Agent Harness Architecture v1.1.0: Least Sufficient Governance
Trajectory control, obligation-first H0–H3 depth, relevance-scoped tools and independently verified deployed effects.
Read article ↗Architecture release · 11 Sep 2026AI Agent Harness Architecture v1.0.0: What Changed
The first formal semantic-version baseline covering decision authority, governed memory, runtime trust, evidence, qualification and continuous assurance.
Read article ↗Harness Weekly
Weekly pieces are cross-layer discovery and evidence briefs. Findings from them only become thematic articles after they map to one architecture anchor.
Harness Weekly: 24 October–1 November 2026 — Evidence Before Effects
Five source-attributed findings on monitoring, identity, skills, security evidence and privacy controls.
Read article ↗Harness Weekly · catch-up prepared 27 Oct 2026Harness Weekly: 19–23 October 2026 — Control Signals
Agent Skills stewardship, named capability telemetry, GitHub Apps, public secret monitoring and PHP remediation.
Read article ↗Harness Weekly · published 18 Oct 2026From Agent Experiments to Governed Systems: Five Engineering Lessons
Governed software delivery, fail-open tests, independent barriers, provenance and repository integrity.
Read article ↗Harness Weekly · published 9 Oct 2026Harness Weekly: 2–9 October 2026
Long-horizon observability, orchestration memory, autonomous evaluation design and harness-aware distillation.
Read article ↗Harness Weekly · published 2 Oct 2026Harness Weekly: 25 September–2 October 2026
Executable benchmark contracts, least-sufficient harnesses, failure-transparent reporting, MCP recovery and budgeted external observations.
Read article ↗Harness Weekly · published 21 Sep 2026Harness Weekly: 15–21 September 2026
Identity, runtime monitoring, misalignment reporting, voice agents, pre-execution auditing and qualified self-improvement.
Read article ↗Harness Weekly · backfill 21 Sep 2026Harness Weekly: 8–14 September 2026
Government harness guidance, tool-relevance controls, runtime hardening, CRA reporting and native execution evidence.
Read article ↗Harness Weekly · backfill 21 Sep 2026Harness Weekly: 1–7 September 2026
Supervised iteration, provenance-aware memory, meaningful human oversight, skill admission and evidence-grounded evaluation.
Read article ↗Separate governance tracks
Adjacent models remain available without being mixed into the Harness layer series.
Decentralization Does Not Remove Responsibility
A practical responsibility model for decentralized systems: factual control, activity, decision authority, economic benefit and assurance evidence.
Read article ↗Blockchain governance · 12 Aug 2026The Blockchain Responsibility Model
Assign capable, accountable owners across protocols, infrastructure, applications, governance, data, and users.
Read article ↗