AI CAPABILITY • IMPLEMENTATION

Agents & Orchestration

From skills to coordinated teams

In 30 Seconds

The paradigm has shifted — twice. First: we stopped building separate agents for each domain. Instead, we build skills — packaged expertise that a general-purpose agent loads on demand.

Now: that agent can coordinate teams. Multiple AI instances working in parallel, sharing findings, dividing complex work — each loading the skills they need. The pattern is evolving from “one agent, infinite expertise” to “coordinated teams with shared expertise.”

Skills remain simple: organised folders containing instructions, scripts, and reference materials. Version them in Git. Share them across teams. Anyone — human or AI — can create them.

Why it matters: Every skill you build compounds. Enterprise knowledge becomes reusable capability — and with agent teams, that expertise scales across parallel workstreams without rebuilding from scratch.

The Paradigm Shift

The field has evolved through three stages — each building on the last.

As Anthropic put it: “After we built Claude Code, we realized that Claude Code is actually a general purpose agent.” And now, that agent can coordinate teams.

Stage 1

Multi-Agent

Separate agent per domain

Each needs its own tools and scaffolding

Coordination complexity grows exponentially

Hard to share learnings across agents

Stage 2 — Established

Skills-Based

One agent + library of skills

Skills loaded at runtime when needed

Expertise compounds across the organisation

Anyone (human or AI) can create new skills

Stage 3 — Emerging

Agent Teams

Lead agent coordinates teammates in parallel

Each teammate loads skills independently

Shared task lists with dependencies

Direct messaging between agents — not just hub-and-spoke

What Are Skills?

Packaged expertise that extends what an agent knows how to do.

Skills are deliberately simple: organised folders containing instructions, examples, and reference materials. Version them in Git. Share them across teams. Anyone – human or AI – can create them.

Terminology note: “Skills” means different things in AI. We follow Anthropic's Agent Skills model – filesystem-based expertise that loads through explicit triggers, not AI-decided retrieval. See AI Agents Explained for full terminology.

Foundational Skills

New capabilities the agent didn't have before. Document creation, data analysis, scientific research – domain expertise packaged for reuse.

Platform Skills

Help agents work better with specific tools. Browser automation, workspace integration, API orchestration – built by platform providers.

Enterprise Skills

Organisation-specific knowledge. Internal processes, coding standards, compliance requirements – institutional knowledge made actionable.

Skills + MCP: The Complete Picture

Skills provide expertise. MCP (Model Context Protocol) provides connectivity. Together, they make agents genuinely capable.

Skills = Expertise

What: Procedural knowledge, best practices, domain expertise

How: Loaded into context when the agent decides to use them

Example: Financial reporting skill, compliance checking skill

MCP = Connectivity

What: Tools and data from external systems

How: Standardised protocol for connecting to outside world

Example: Database queries, API calls, file system access

Skills often orchestrate multiple MCP tools – they provide the “how” while MCP provides the “with what.”

The Computing Analogy

Skills are to AI what applications are to computers

Applications = Skills

Domain expertise, unique capabilities

Operating System = Agent Runtime

Context management, tool execution

Processor = Model

Intelligence, reasoning capability

A few companies build processors and operating systems. Millions build applications. Skills open this layer to everyone.

Orchestration Patterns

How skills and subagents coordinate to get work done

Chaining

Sequential handoffs

Task flows through skills in sequence. Research skill feeds analysis skill feeds reporting skill. Best for linear workflows with clear stages.

Routing

Smart direction

Classify requests and load the right skill. Customer support, technical queries, sales questions – each triggers the relevant expertise.

Parallelisation

Simultaneous work

Spawn subagents to tackle different aspects at once. Research, analysis, and drafting happen simultaneously, then results are synthesised.

Orchestrator-Workers

Dynamic delegation

The main agent breaks down tasks and spawns workers with relevant skills. Flexible – subtasks aren't predetermined, but discovered.

Agent Teams

Coordinated parallel work

NEW

A lead agent coordinates multiple teammates working in parallel. Teammates communicate directly via messaging, share a task list with dependencies, and each loads skills independently. Genuinely parallel work with convergence — not just delegation and reporting back.

Evaluator-Optimizer

Iterative refinement

Generate with one skill, evaluate with another. Loop until quality threshold is met. Like a writer and editor working together.

Human + Agent Collaboration

Agents don't replace humans. They extend human capability. The question is: how much autonomy, and where does human judgment stay essential?

The Autonomy Spectrum

Human in the Loop

Human approves each step. Maximum oversight. Best for high-stakes decisions.

Human on the Loop

Agents work, human monitors. Intervention when needed. Best for routine work with exceptions.

Human out of the Loop

Full automation with guardrails. Human review of outcomes. Best for well-defined, low-risk processes.

Most organisations in 2025-26 will operate “in the loop” or “on the loop”

Success depends on knowing where on this spectrum each task belongs.

EMERGING

Networked Agents: A Different Pattern

Beyond agent teams — which operate within a single session under human oversight — a different pattern is emerging: AI agents communicating with other agents across boundaries. Early experiments show agents forming communities, coordinating autonomously, and exhibiting emergent behaviour at scale.

Skills-Based

Established

  • • One agent + library of skills
  • • Human in/on the loop
  • • Contained context, auditable
  • • Expertise compounds internally

Agent Teams

Early adoption

  • • Lead + teammates in one session
  • • Human oversight of the team
  • • Shared task list, direct messaging
  • • Parallel work with convergence

Networked Agents

Experimental

  • • Many agents interacting peer-to-peer
  • • Autonomous coordination
  • • Emergent behaviour, harder to predict
  • • New security surface (agent-to-agent)

Why This Matters

Networked agents raise questions we're only beginning to understand:

  • Security: Prompt injection between agents, credential exposure, context leakage
  • Governance: Who's responsible when agents coordinate autonomously?
  • Predictability: Emergent behaviour is fascinating but hard to audit

We track these developments closely. For most organisations today, skills-based + agent teams offers real capability with governance. Networked agents are fascinating but require governance models that don't yet exist.

Current example: OpenClaw agents interacting via Moltbook demonstrated community formation, coordination protocols, and emergent capabilities - all without human direction. See our AI Landscape for tool coverage.

Three Principles We Follow

From Anthropic's guidance on building effective agents

1

Simplicity

Start with the simplest solution that works. Add complexity only when it demonstrably improves outcomes.

2

Transparency

Show the agent's planning and reasoning. Humans should understand what's happening and why.

3

Tool Design

Agent-Computer Interfaces need as much care as Human-Computer Interfaces. Well-designed tools make agents reliable.

Expertise That Compounds

Whether you're exploring how skills could extend your AI capability, designing your organisation's knowledge architecture, or need hands-on implementation – we can help build expertise that compounds.