Two weeks ago, I told you the role card is your new org chart. Outputs. Split. Harness. The first two are conceptual. The third is where everyone stalls — because the moment "harness" enters the conversation, every non-engineer in the room starts looking for the exit.

Don't. The harness is the easy part. It just doesn't look like it.

Here's the data forcing the issue. IDC says AI agents will be embedded in 80% of enterprise applications by the end of 2026, up from less than 15% in 2024. Gartner forecasts 40% of new enterprise software will be built through "vibe coding" — natural-language prompts producing working logic — within the same window. 35% of organizations report broad agent usage today; 17% have rolled them out company-wide. Salesforce calls it a 282% jump in adoption in twelve months.

That's the headline. Here's the buried lede: Gartner also says 40% of those agentic AI projects will be cancelled by 2027 because of messy governance, weak permissions, and unclear ROI.

Forty percent. Not because the models are bad. Not because the no-code platforms are bad. Because the people deploying agents are skipping the harness and going straight to the agent. They're hiring an employee with no job description, no manager, no expense limits, and no performance review — and then acting surprised when something breaks.

The harness has four parts: memory, tools, guardrails, evaluation. Most non-engineers build the first two and skip the last two. That gap is where the dead 40% lives.

The first two are obvious. Memory is where the agent stores what it needs to remember — past context, decisions made, exceptions logged. Build it in Airtable, Notion, Sheets. Tools are the systems the agent can read from and write to — email, calendar, CRM, payments. Wire them up in Zapier, Lindy, Make, Relay. Every modern no-code platform ships with 200+ pre-built integrations.

The other two are where projects die. Guardrails are the agent's expense policy — what it can't do, what it needs approval for, what it can do alone. Evaluation is the agent's performance review — how you know it's not just running, but right.

Skip those, and your agent is an employee with no manager. Build them, and you have an orchestrator's instrument instead of a liability with API access.

This is what Maestría buys you that a license to a no-code tool doesn't. The schema only you can design. The guardrails only you know to write. The evaluation rubric only you can score.

The harness is what you know about your business that no platform can sell you.

Build in this order. Skip a step and you join the failed 40%.

1. Memory. Structured records the agent reads from and writes to. Not the model's context window — that resets every conversation. The part that persists. Build in Airtable, Notion, or Sheets. The orchestrator's move: design the schema before connecting the agent. If you can't draw the table on paper in 15 minutes, the agent doesn't have enough structure to remember anything useful.

2. Tools. The systems your agent can touch. Email, CRM, calendar, files, payments, the database from step 1. Connect them in Zapier, Lindy, Relay, Make, or Microsoft Copilot Studio. The orchestrator's move: minimum viable toolkit. Read access first, write access second, payment access last. Each tool is a capability and a liability.

3. Guardrails. Three lists on a page. Hard nos — things the agent literally cannot do, ever. Approval-required — things the agent proposes, a human approves. Auto-execute — things the agent does on its own. Wire approvals through Slack or email. Most agents should start with 80% in approval-required and earn their way down. A spending agent that can't spend more than $50 without approval is a useful agent. A spending agent with no cap is a lawsuit waiting for a calendar invite.

4. Evaluation. A weekly 20-minute calendar block. A 5-output spot-check rubric scored 1–5 on accuracy, tone, judgment. A side-by-side test against how you would have done it. If there's no calendar block, there is no evaluation, and the agent is running blind.

IDC: 80% of enterprise apps will have embedded agents by end of 2026. Up from less than 15% in 2024. Pair with Salesforce's 282% adoption jump and the curve is clear — this is no longer early adopters, this is the middle of the wave. → https://www.salesmate.io/blog/future-of-ai-agents/

Gartner: 40% of agentic AI projects will be cancelled by 2027. Cause of death: governance, permissions, unclear ROI. The model isn't the problem. The harness is. → https://www.aninews.in/news/business/top-7-agentic-ai-programs-to-build-autonomous-ai-systems-in-202620260501125741/

Vibe coding hits the enterprise. 40% of new enterprise software will be built via natural-language AI generation by end of 2026 — Airtable Omni, Microsoft Copilot Studio, Joget, Replit Agent. The layer that used to require an engineer is now a prompt window. → https://www.airtable.com/articles/no-code-ai-tools

Identity is the new product surface. May 2026 reporting from Okta and government cyber agencies: agents inheriting full human permissions are the next enterprise breach vector. Vendor whitelists and scoped credentials belong in the guardrail layer, not retrofitted later. → https://blog.mean.ceo/ai-agents-news-may-2026/

Vertical agents are beating general-purpose ones. Narrow-scope agents — lead research, claims processing, contract redlines — are outperforming "do everything" assistants on every measurable KPI. The role card discipline from Issue 003 isn't aesthetic. It's the entire reason the agent works. → https://blog.mean.ceo/ai-agents-news-may-2026/

This week: Build your first Agent Org Chart.

Take the framework from The Pattern and apply it to one function in your business. Pick the area where you or your team waste the most time on repetitive work.

Draw a simple org chart with three boxes:

Write one paragraph describing how work flows between the three boxes on a typical day.

You just designed your first agentic workflow — not by learning to code, but by learning to manage.

That's the skill that separates the next generation of operators from the last.

Agentic Congruence is a weekly newsletter about orchestrating ventures, agents, and systems. Published by La Maestría. Reply to this email anytime — I read everything.