---
title: "Subagent Architecture: Why Specialised Agents Beat One Generalist"
url: https://xalt-northstar.com/blog/subagent-architecture/
description: "Specialised subagents with skills, a clear orchestrator and full logging: what we learned building agent systems in production, and where governance fits in."
source: XALT Northstar
---

[Blog](https://xalt-northstar.com/blog/) Agent architecture

# Subagent Architecture: Why Specialised Agents Beat One Generalist

Specialised subagents with skills, a clear orchestrator and full logging: what we learned building agent systems in production, and where governance fits in.

Philipp Göllner 6 October 2026 4 min read

A general AI assistant is acceptable at many things. A set of specialised subagents, each with its own context, tools and instructions, is good at one thing each. Put an orchestrator in front that calls the right specialist, and the system gets better results than one assistant that tries to do everything.

The idea is now popular in engineering circles. For enterprises, the interesting part is not the number of agents. It is the architecture that keeps them reliable and auditable. This post describes what we have built and what we learned from it.

## What a subagent architecture is

Three building blocks show up in every system we build:

- **Skills.** A skill is a versioned package of instructions, tool bindings, conventions and examples, for example a `SKILL.md` for working with Jira or Slack. The agent loads only the skills a task needs.
- **Subagents.** A subagent is a specialist with its own context window and a narrow job: transcribe, classify, write tickets, review code.
- **An orchestrator.** The orchestrator splits the work, calls the right subagent and collects the results.

The separation matters because context is the scarce resource. A subagent that sees only what its job requires makes fewer mistakes than a generalist whose context is full of everything else.

## Examples from our projects

**A self-hosted agent with a skill system.** We run a personal agent on the open-source project OpenClaw. Instead of one large prompt, it uses versioned skills for Atlassian, Slack, GitHub, messaging, transcription and health checks. At runtime the orchestrator decides which skills to load. Specialisation becomes something you can share, version and audit.

**An internal operations agent.** Our operations agent lives in Slack and has access to Jira, Confluence and our CRM. When someone asks for the status of a project, it answers fast because it has the right context, not because its model is larger.

**Voice to tickets on construction sites.** A site manager records while walking the site. One subagent transcribes the audio. A second recognises defects and status changes. A third turns them into Jira tasks with defect categories and owners. Three specialists, one workflow, and nobody types tickets by hand.

**Several agents on one airfield.** For airport operators, separate agents handle safety self-inspections, detection of foreign object debris on the runway and wildlife management. Each has its own model, its own training domain and its own compliance logic under EASA rules.

## Lessons from running subagents in production

### Skills beat prompts

A prompt is a note. A skill is infrastructure: versioned, documented, bound to tools and tested. Teams that still copy prompts from a document have a habit, not an architecture.

### Orchestration and logging are required

Many subagents that know nothing about each other create confusion, not capability. We build every subagent with two properties: a clear hand-over (what does it return, and to whom?) and logging (why did it decide this?). Without both, debugging becomes guesswork, and an audit becomes impossible.

### Domain context matters more than model size

The aviation agents work because the regulations, airport operating standards and wildlife patterns travel with them as context. The real work is curating domain knowledge so that a subagent applies it reliably.

## Governance grows with the number of agents

The more agents you run, the more pressing a plain question becomes: which agent identity has which permissions in which system? A subagent architecture without governance multiplies risk with every new specialist.

The controls are the same as for a single agent, applied to each one:

- **Approval for every action** that changes a system of record or leaves the company.
- **A complete activity log** across orchestrator and subagents, so you can trace a result back to the agent and the decision that produced it.
- **An inventory of agent identities** and their permissions, reviewed on a schedule.
- **Freedom to choose the model per subagent**, so a specialist can use the provider that fits its task.

## What this means for your roadmap

- One general assistant for everyone is a starting point, not an architecture.
- Many agents without governance turn into an audit problem.
- Plan subagents, skill versioning and governance together. That combination is what scales.

**Related:** [Agentic Engineering](https://xalt-northstar.com/agentic-engineering/) – From developer self-service to an agentic platform. · [Regulated AI](https://xalt-northstar.com/regulated-ai/) – Operational AI for the regulated enterprise.

Your next step

## Put this to work in your team.

Start with something you can use today, or work on it hands-on with us. No sales call needed.

- **Run the Agentic Coding Workshop on your repo** Four hours, remote, on your own repository: the workflow set up with templates your team keeps. [See the workshop](https://xalt-northstar.com/agentic-engineering/#agentic-coding-workshop)
- **Fix a real problem live in 90 minutes** Bring one agent task that is too broad, and split it into specialised agents with their own permissions. [Book the live session](https://xalt-northstar.com/agentic-engineering/#live-fix-session)

Read next

- [Personal agents Personal AI Agents in the Enterprise: Pick the Type Before the Tool 6 October 2026 · 5 min read](https://xalt-northstar.com/blog/personal-ai-agents-governed/)
- [Cost & models Multi-Provider AI: The Right Model per Task, With Costs Under Control 6 October 2026 · 5 min read](https://xalt-northstar.com/blog/multi-provider-ai-model-per-task/)

---

Solutions

- [Agentic Engineering](https://xalt-northstar.com/agentic-engineering/)
- [Regulated AI](https://xalt-northstar.com/regulated-ai/)
- [AI Services](https://xalt-northstar.com/ai-services/)
- [AI Brain](https://xalt-northstar.com/brain/)
- [Visual Review](https://xalt-northstar.com/visual-review/)
- [Personal Agent](https://xalt-northstar.com/personal-agent/)

More

- [Events](https://xalt-northstar.com/events/)
- [Blog](https://xalt-northstar.com/blog/)
- [Careers](https://xalt-northstar.com/careers/)
- [Contact](https://xalt-northstar.com/contact/)

Company

- [xalt.de](https://www.xalt.de/)
- [Imprint](https://xalt-northstar.com/imprint/)
- [Privacy](https://xalt-northstar.com/privacy/)

**XALT** Northstar

In progress SOC 2 Type 1 · ISO 27001 · ISO 9001 · GDPR. Not certified yet. [Trust Center](https://trust.inc/xalt)

© 2026 XALT LLC
