Most companies use AI as isolated tools. A chat here, a copilot there, an automation somewhere else. Tools that don't talk to each other, that don't remember what they did yesterday, that don't know what the colleague next to them is doing. This isn't using AI — it's collecting AI.
What separates casual use from operational use is the harness — the infrastructure that connects, orchestrates and governs AI agents into a coherent system. According to Mitchell Hashimoto, creator of Terraform, the formula is Agent = Model + Harness. The model is the brain. The harness is everything else. At Capiva, we built an operational harness. And it runs in production every day.
What is an AI Harness (and why the model alone isn't enough)
Mitchell Hashimoto — creator of Terraform and one of the most respected infrastructure engineers in the industry — formalized the concept in February 2026: Agent = Model + Harness. The model is the brain. The harness is everything else: connected tools, persistent memory, execution rules, guardrails, feedback loops and observability.
Martin Fowler and Birgitta Böckeler from Thoughtworks expanded with the guides and sensors taxonomy. Guides are rules that direct the agent — what to do, what not to do. Sensors are mechanisms that detect when something goes off track. According to research published in the Thoughtworks Technology Radar, every robust harness needs both.
According to Deloitte's State of AI in the Enterprise report (2026), 88% of companies use AI but only 29% achieve real ROI. The gap is exactly the harness. Companies have models. They don't have the system around them.
The architecture that runs at Capiva
Each component exists for a specific operational reason. Nothing was added out of technical curiosity. These are 8 components integrated into an operational harness.
Claude Code CLI as primary interface
Claude Code is a command-line interface that integrates directly with the filesystem, version control and automation. According to Anthropic, the agent operates inside the work environment — reads files, edits code, executes commands and creates artifacts. The barrier between asking AI and AI executing disappears, resulting in productivity gains of 3 to 5 times in engineering tasks.
Obsidian as structured knowledge base
The vault contains over 1,600 markdown documents organized by domain, with YAML metadata and content maps for conceptual navigation. This solves a problem that McKinsey research (2024) identifies: professionals spend 19% of their time searching for information scattered across different tools. In the vault, everything is in one place, version-controlled by git and accessible by AI.
Semantic search (QMD)
QMD combines three approaches: BM25 for keywords, vector embeddings for semantics and LLM re-ranking. According to research published on Retrieval-Augmented Generation (Lewis et al., 2020), combining dense and sparse search improves retrieval quality by up to 37%. The result: when the agent needs context, it finds the most relevant document among 1,600+ options in seconds.
MCP: the integration layer
Model Context Protocol is the open standard — now under the Linux Foundation — that allows the agent to use external tools natively. In practice, Claude connects directly to n8n for workflow automation, Chrome DevTools for testing, Granola for meeting transcripts and Context7 for technical documentation. Hashimoto's analogy is USB-C: a standard that lets you plug in any tool without rewriting integrations.
Autonomous agents (Boss/Worker)
The harness doesn't depend on constant human interaction. Boss supervises, prioritizes and delegates. Worker executes board tasks in regular cycles without human intervention. Applied audits the system itself, identifies improvements and implements one per day. The principle: what can be automated shouldn't consume human attention. What requires human judgment receives full human attention.
Persistent memory across sessions
Every work session produces context. In most setups, that context is lost when the window closes. In the harness, it persists across multiple layers: working memory for session state, auto memory for accumulated preferences and patterns, and vault for permanent knowledge. The equivalent of an employee who never forgets what was discussed in previous meetings.
Skills as replicable workflows
Instead of writing long prompts every time a recurring task is needed, skills encode the entire process: objective, steps, output template, constraints. A single command triggers the whole workflow. Skills for research, idea capture, content creation and transcript analysis — each one is an SOP executable by AI.
n8n for workflow automation
Workflows involving multiple systems — emails, webhooks, APIs, databases — run on n8n. According to n8n (2026), the platform processes over 150 million executions per month across its user base. The MCP server connects the agent to n8n, enabling it to trigger, monitor and consume automation results within the same harness.
Guides and sensors: harness governance
Fowler and Böckeler's taxonomy applies directly to the operational harness. Guides direct behavior: vault operation rules, approval protocol (when to ask, when to execute), quality of thought (verify before building) and board-first (all work goes through the task board before execution). Sensors detect deviations: automated daily audit, task board verification, quality gates between phases and feedback hooks that capture permanent corrections.
Hashimoto's principle in action: every error becomes a fix in the harness so it never repeats. The system improves with every cycle. According to research on continuous improvement published by IEEE (2025), systems with automated feedback loops reduce error recurrence rates by 68% compared to manual retrospective processes.
The operational result
A founder operating with an AI Harness produces output equivalent to a traditional consulting team. Simultaneous projects in Brazil, US and UK. Multiple products built and maintained in parallel. Content pipeline, diagnostic tools and client work — all running simultaneously with a single dedicated resource.
In an implementation as AI Center of Excellence for a global Fortune 500 consumer goods company, this same approach compressed corporate project cycles from 6 months to 2 weeks — a 92% reduction in delivery time. The model replaced teams of 10 people with a single resource operating in short sprints of discovery, prototyping and deployment. The delivery velocity became an internal benchmark for other teams in the organization.
What this means for your company
AI Harness is not a product you buy. It's an architecture you build. Every company needs a different harness because every operation has different tools, processes and constraints. The components are accessible — Claude Code, Obsidian, n8n, MCP — all available. The challenge is not access to tools.
The challenge is knowing how to connect them into a system that works, that self-improves and that scales without adding people. That's the work Capiva does. We design and implement AI Harnesses for operations that want to go from "we use AI" to "AI operates our company." According to Deloitte, the difference between the 29% that achieve ROI and the 71% that don't is in the infrastructure, not the model.
Want to build an AI Harness for your operation?
Capiva designs and implements custom AI Harnesses. The first step is a Strategic Diagnosis that maps where AI creates value in your operation.
Talk about AI HarnessHarness engineering for businesses: what you need to know
Harness engineering is the discipline of designing complete infrastructure around AI agents for reliable production operation, formalized by Mitchell Hashimoto in February 2026 with the formula Agent equals Model plus Harness. According to Deloitte's State of AI in the Enterprise report (2026), 88% of companies use AI but only 29% achieve real return — the gap is the harness infrastructure. The Capiva Agent OS is the proprietary operational harness implementing this architecture with dozens of autonomous agents executing approximately 500 task cycles per month; specifically, a self-improvement subsystem logs operational decisions, detects patterns and implements system corrections autonomously — each error permanently encoded so it never recurs. Additionally, Martin Fowler and Birgitta Böckeler from Thoughtworks expanded the discipline with guides (rules directing behavior) and sensors (deviation detection), both of which the Capiva Agent OS implements through codified governance and automated daily audits. The harness has been running in continuous production for over 3 months, integrating tools via MCP (Model Context Protocol), persistent memory across sessions and workflow skills with permanent feedback loops. In emerging markets, harness engineering adoption is practically nonexistent, representing a significant competitive window for organizations that build their operational AI infrastructure first.