Post

Skills, Connectors, and Subagents: Decoding the Architecture Anthropic Just Made Standard

May 9, 2026· 3 min read

Originally published on LinkedIn (May 9, 2026). Republished here with light copy-editing.

Anthropic released a production architecture for agentic AI in regulated industries on May 5th, including ten agent templates for financial services, Microsoft 365 add-ins, eight data connectors, and a Moody’s MCP app. The significant insight lies not in customer logos but in the underlying orchestration topology.

The reference architecture: skills, connectors, subagents

The foundational principle: “Each agent template is a reference architecture that packages three things: skills (instructions and domain knowledge for the task), connectors (governed access to the data the task runs on), and subagents (additional Claude models called upon by the main agent, for specific sub-tasks).”

Skills: scoped instruction modules with lazy loading

A skill is a structured instruction package, a SKILL.md plus supporting assets like templates, schemas, exemplars, and tool specs, loaded conditionally based on task classification. This matters for context window economics.

Rather than embedding all methodology, formatting rules, comparables logic, and review criteria in a monolithic system prompt, lazy-loading skills reduces per-call input token cost significantly. A 40K-token always-loaded prompt versus an 8K controller plus lazy-loaded 6K skill represents approximately 3x the input cost reduction per call, compounding across multi-turn workflows.

A typical skill structure includes:

• SKILL.md (entry-point instructions)
• templates/ (cover_page.pptx, comparables_table.xlsx)
• schemas/ (target_company.json, comparable_set.json)
• exemplars/ (good_pitch_001.md, bad_pitch_anti_pattern.md)
• tools.yaml (tool subset this skill needs)

The controller’s role is task classification to skill resolution to context injection, not solving the task itself.

Connectors: governed data access with real-time semantics

A connector is not merely an API wrapper. It is a governed, audited, permissioned access layer with three production-critical properties:

1. Per-tool authorization: agents receive scoped tokens, not blanket access
2. Audit-grade logging: every call recorded with parameters, timestamp, and caller identity
3. Schema-aware reasoning: connectors expose typed responses for planning

This differs structurally from RAG-over-PDFs approaches. A pitchbook generation task flows through:

1. Controller receives target list and classifies it as “comparables-driven pitch”
2. Loads skill: pitchbook_builder
3. Calls FactSet connector for fundamentals on the target
4. Calls PitchBook connector for the private comparable set
5. Calls Capital IQ connector for multiples on public comps
6. Calls Guidepoint connector for expert interview excerpts on the sector
7. Spawns subagent: comparables_selector
8. Spawns subagent: methodology_validator (LLM-as-judge pattern)
9. Generates draft, then human-in-the-loop review and approval

Every connector call is logged; every subagent decision is traceable for compliance replay.

Subagents: specialized inference with independent context

A subagent is a separate Claude invocation with its own system prompt, tool subset, and context window, called by the controller for a narrow sub-task. Critically, it does not share the parent’s full context. It receives only what is relevant.

Key technical advantages:

• Context isolation prevents methodology bleed across sub-tasks
• Specialized prompting lets each subagent run with task-tuned instructions
• Independent failure modes mean a failed check fails that subagent only; the controller can retry, escalate, or route to a human
• Granular observability allows independent instrumentation in monitoring tools

Example architecture from my own document pipeline:

• Controller (orchestration only, no domain reasoning)
  • Subagent 1: custom CNN for page classification
  • Subagent 2: Claude Opus 4.5 for vision-based OCR to markdown
  • Subagent 3: AWS Nova 2 Pro for schema translation
  • Subagent 4: DeepSeek for automated prompt optimization
  • Subagent 5: LLM-as-judge for output validation

Where this pattern bites you

Every architecture has tradeoffs:

• Latency compounds: five sequential subagent calls mean five inference round-trips. Without parallelizing independent calls, a “fast” system becomes a 40-second wait per task.
• Debugging gets harder: a failure in subagent 4 that depends on outputs from subagents 1 to 3 requires deterministic replay of the full chain. Audit logging alone is insufficient.
• Cost can rise, not fall: overlapping context naively re-sent to each subagent multiplies input token cost by the number of subagents. Context engineering, deciding exactly what flows into each call, separates clean decomposition from expensive overhead.

Closing thought

The ten templates will make headlines. The real signal is architectural. Skills, connectors, and subagents is now the canonical production pattern for agentic AI in regulated industries. Teams that win the next 24 months will have the cleanest decomposition, not the smartest model.

model-analysisagentic-aiarchitectureregulated-industries