Introducing HAMMER, the Homegrown Agent Maturity Model for Execution & Risk

The Homegrown Agent Maturity Model for Execution & Risk (HAMMER): is a five-level framework with seven domains for security teams to locate where they stand on securing homegrown agents, then make the case for what they need to advance.

The most powerful AI agents in your enterprise may be ones you never bought. Your own teams are building them by wiring models, tools, and data into software that acts on its own, often with little security approval or oversight.

Why Homegrown Agents Need Their Own Yardstick

The threats are the same as for any agent, but the difference is who owns the gaps. Platforms ship some security features out of the box, but the real risk with a homegrown agent is what those features don't cover, such as gaps inside what a platform provides, or assumptions that something is covered when it isn't.

Homegrown agents don’t stand still. New use cases, tools, triggers, and sub-agents keep changing what homegrown agents can do long after the launch review, shifting its execution path and whether your controls still apply. Since a homegrown agent is a moving target, HAMMER scores how well you secure it as it evolves, not just on the day it ships.

How the Model Works

The seven domains describe what to assess across your homegrown agents. Organizations are uneven, so each domain is scored independently. But real exposure rarely sits inside a single domain. It sits in the execution layer that runs across all of them, acting as a bridge across the capability domains, which is why HAMMER scores it as a domain of its own. The goal isn't to reach Level 5 across every domain ,it’s evidence-backed progress you can repeat as agents change.

Six Capability Domains

These six capability domains, also known as the pillars, provide the foundation for securing the core components of homegrown agents:

1. Discovery & Inventory. Which agents exist, who owns them, what do they touch, and how do they start?

2. Identity, Access, & Secrets. Does each agent have its own scoped, least-privilege identity with managed secrets?

3. Secure Development & Design. Is security built into how agents are designed, tested, and shipped, rather than "vibe-coded" from a template?

4. Runtime Protection & Guardrails. What can an agent actually reach, and what passes through it?

5. Detection, Monitoring, & Response. Can you see agent behavior, attribute it, detect harm, and then contain it?

6. Governance, Metrics, & Ownership. Who owns each agent, which policy applies, and do the metrics show it enables the business?

Domain 7: The Bridge Across the Pillars

Risk doesn't stay inside the six capability domains. It runs across them, in the execution sequence. An agent is invoked, reasons, calls tools under some credential, touches data, hands work to other agents, and then acts. Each step can be logged, permitted, and authorized while the whole sequence is harmful, invisible.

Domain seven, Execution Visibility & Risk, is horizontal because it scores three things across your controls:

1. Can you surface what agents do?
2. Can you enforce policy on the whole sequence?
3. Can you protect against harm no single control sees, even as agents take on new use cases and flows?

It matters most for agents delegating tasks, connectors and skills that change monthly, models swapped under a live agent, new security features from stack providers, and even agent-to-agent communication, where trust, permissions, and attribution cross boundaries no single pillar owns. Strong controls with weak cross-control visibility can leave organizations exposed. We expect most organizations to score lowest in this domain. That gap tracks real-world risk more closely than any single pillar.

Self-Assessment Prompts: Turn Them Into Your Own Agentic Skills

We're releasing HAMMER's self-assessment as three prompts for the LLM your team already uses. No new tool or budget needed. Agents evolve, so only conducting this one time will leave you with stale information. They're prompts, not a hosted tool, so you can convert them into agentic skills, then run them on a recurring basis.

What the three prompts do:

• Evaluate. An adaptive interview of up to 30 questions, that you’ll work through one at a time. It calibrates for agent autonomy, separates what you claim from what you can prove with hard data, and then returns a score for each of the seven domains.
• Plan. Builds a phased, prioritized action plan, including concrete asks per owner (engineering, security, product, DevOps, leadership), and an editable timeline.
• Brief. Puts together a one-page executive summary that includes the ask, the current risk, the cost, the return, and the decision requested.

To turn them into a skill, package HAMMER as a custom GPT, Claude Project, or internal skill. Simply attach the framework as knowledge, use Prompt 1 as standing instructions, save Prompts 2 and 3 as follow-ups, add read-only tools for your inventory, architecture docs, and log every run so results stay comparable. Reassess quarterly, or whenever an agent gains a new capability.

Get the Full Framework

This article is just the high-level view. The full document includes the complete matrix for every domain, the five maturity levels in detail, scenarios for using this framework in practice, mapping to industry frameworks (including the OWASP Top 10 for Agentic Applications), and all three prompts for you to use.

Download the full HAMMER framework to get started.