Strattum Skills Every enterprise has repeating procedures — reports, queries, approval rules. Skills transforms each one into a tool any agent can invoke, respecting who can call what.
name: customer-health inputs: account_id: string risk_level: read-only owner: data-platform Calculate the health score for account {{account_id}}: cross NPS, open P1 tickets, overdue payments and usage over the last 30 days. SELECT health_score, top_signals FROM customer_health WHERE account_id = {{account_id}}
Every new agentic tool competes for space on the engineering roadmap. Skills removes that competition: the same analyst who writes the weekly query writes the Markdown — engineering reviews in a standard PR, and a minute later any company agent can use it.
customer-health.md using the query that already runs every Monday. A Skill is a single .md file with three parts: identity, instruction, and execution. Each part serves a purpose — no code, prompts, and documentation scattered across different places.
Who it is, who owns it, what it can do. This is the first part Strattum reads on every call — before the agent even acts. It applies permissions and records every invocation.
In natural language, you write the Skill's objective and business rules. This text becomes the instruction the agent follows when the Skill is called. Rule changed? Edit the text — no code rewrite.
When it needs to query data or run logic, you embed SQL or Python directly in the file. Code runs in an isolated, secure environment with the caller's permission — and the result returns as context for the agent to respond.
Skills do not outsource governance to the system prompt. Every execution passes through Strattum's runtime, which applies permission, approval, quota, and audit before the LLM sees the result. The CISO does not need a second model to review.
The Skill executes with the credentials of the user who called the agent. Salesforce, SharePoint, core banking, and SAP ACLs remain in effect inside the prompt. RBAC + ABAC applied at runtime — no second authorization model.
Skills marked as mutable or sensitive trigger approval before execution. Configurable webhook for Slack, Microsoft Teams, email, or your ITSM. Approval is logged with identity and call context.
Every execution runs in an isolated microVM — ephemeral filesystem, explicitly declared network, CPU and memory limits applied. No risk of a Skill compromising the cluster or accessing data outside its declared scope.
Every call logged with identity, parameters, queried source, and approvals. Structured export for Splunk, Datadog, or your SIEM. Ready for BCB 4658, LGPD, and internal audit without additional work.
From fintech to manufacturing — examples of how Skills apply across distinct domains. Your business's Skills are built by your team, with support from Strattum's FDE, all following the same standard: declared parameters, expected-format return, and the same governance as Strattum Data Foundation.
Health score per account crossing NPS, tickets, payments, and usage.
account_id: "acme-corp" Contracts nearing renewal with churn signals estimating exposed revenue.
window: "next-90-days" Unified contact profile — activity, deals, support, knowledge.
email: "ana@client.com" Failure probability per equipment crossing MES, IoT, and history.
machine_id: "EXT-204" Aggregate exposure by sector with BCB 4658 rules and auditable standardization.
sector: "agro" · window: "Q3" Upsell opportunities ranked by usage signal and product fit.
segment: "mid-market" Ticket summary by company, priority, and average response time.
company: "globex" · period: "7d" Portfolio performance summary with revenue delta and risk flags.
fund: "growth-2024" Representative catalog examples. Skills proprietary to your domain are developed by your team or by Strattum's FDE.
Skills is not the first attempt to give agentic tooling to an organization. Worth comparing with what is in production today — without trash-talk, with honesty about when each option makes sense.
| Strattum Skills | LangChain Tools | Custom GPTs | Internal scripts | |
|---|---|---|---|---|
| Authoring language | Declarative Markdown | Python code | ChatGPT UI | Varies by team |
| Versioning | Git-native · PR review | Git (manual) | No history | Git (fragmented) |
| Source permissions | Inherited at runtime | DIY | Not supported | DIY |
| Isolated sandbox | VM-grade, built-in | DIY | Hosted (OpenAI) | No isolation |
| Audit trail | SIEM-ready built-in | DIY | OpenAI-only | No audit |
| Model compatibility | Any MCP client | Multiple models | GPT only | N/A |
LangChain shines when the entire team is software engineering. Custom GPT works for small teams within the OpenAI ecosystem. Internal scripts are reality — but they become operational debt. Skills targets the enterprise that needs governance + model breadth + authoring outside engineering.
The same Skill works in any MCP client. Switching models tomorrow does not require rewriting agentic tools.
Claude 
ChatGPT 
Copilot 
Gemini Distributed via open MCP standard (Anthropic). Strattum does not lock you into a model or client.
In a guided session with the FDE, we show how an SQL query or notebook your team already uses becomes a governed Skill — with inherited permissions, audit trail, and available in any MCP client.