AI Agents & Agentic Workflows
The difference between an automation and an agent is the loop: act, evaluate, iterate until the output passes. This playbook covers the 2026 platform landscape, the build sequence, the architecture patterns that survive production, and what to spend (and charge) for agentic systems.
Format · Playbook · 6 sections · Free, no gate
01The Platform Landscape
| Platform | Best for | Limitations |
|---|---|---|
| Claude Code | Production agent builds, skills, subagents | $20-200/mo, rate limits |
| Claude Routines / Managed Agents | Scheduled and triggered automations hosted by Anthropic | Newer surface, limited networking |
| n8n | Visual workflow builder with an AI agent node; client-facing visual workflows | Being squeezed by natively agentic tools |
| Make.com | Simple trigger-action automations | Not truly agentic (no loops or self-correction) |
| OpenAI AgentKit | Hosted agents with built-in eval tooling | Less flexible than code-first stacks for custom work |
| Local agent OS (OpenClaw-style) | Always-on personal assistant on your own hardware | Early, overhyped for production client work |
The durable skill is not any one platform. It is understanding what a business needs, scoping a solution, and delivering results; the tools keep getting abstracted away underneath you.
02The Build Sequence
Set up the coding agent
Claude Code in your editor of choice. This is the orchestrator everything else hangs off.
Write the project brain (CLAUDE.md)
One file describing project structure, conventions, and constraints. The single highest-leverage artifact in the whole stack: every session starts with full context instead of re-explaining.
Build skills
Reusable prompt templates in .claude/skills/, each doing exactly one thing well. Skills turn a general agent into a fleet of specialists.
Connect MCP servers
MCP is the integration layer: scrapers, email senders, Slack, databases, any API. Connect only what the workflow actually needs.
Create a small agent team
2-4 agents with distinct roles (researcher, builder, QA reviewer) coordinated by the main session. Resist large swarms; coordination overhead grows faster than throughput.
Add quality loops
A QA agent reviews output against explicit criteria and sends it back until it passes. The loop, not the model, is what makes output production-grade.
03Architecture Patterns That Survive Production
Workflow vs agent
A workflow is trigger, action, done. An agent loops: act, evaluate the result, iterate until the output meets quality criteria. Use workflows for deterministic tasks, agents where judgment or recovery is required.
Directive / orchestration / execution
Separate what the agent should accomplish (directive), how sub-tasks are coordinated (orchestration), and the actual tool calls (execution). Debugging gets dramatically easier when these layers are explicit.
Multi-agent team
An orchestrator session plus a researcher (search, data gathering), a builder (code, content, assets), and a QA reviewer. All teammates share MCP servers, skills, and files.
Self-improving loops
Run a skill, detect failure, research the fix, apply the patch, re-run. Skills that work roughly 70% out of the box approach 95% reliability after a few of these cycles. The same loop applies to cold email copy: A/B variants overnight, keep winners, discard losers.
04Cost and Token Optimization
| Method | Cost | Output quality |
|---|---|---|
| Claude Code (frontier model) | $20-200/mo | Reference quality |
| Claude Code via OpenRouter routing | $5-25/mo | 80-90% |
| DeepSeek-class models via OpenRouter | Near zero | ~80% for coding tasks |
| Local models (Ollama) | $0 | 60-70% |
- Route by difficulty: small model for simple tasks, mid-tier for medium, frontier only for complex reasoning.
- Cache aggressively: repeated context should never be paid for twice.
- Use subagents for parallel work; each gets its own context window instead of bloating one session.
- Measure cost per completed workflow, not cost per token: a cheaper model that needs three retries is not cheaper.
05What Agentic Systems Are Worth
Market rates for delivered agent systems (externally validated against 2026 agency pricing surveys and Upwork data). Useful in both directions: what to budget if you are buying, what to charge if you are building.
AI business audit
$5K-15KAgent analyzes operations, identifies automation opportunities, produces a roadmap
Agentic sales pipeline
$3K-10KAgent qualifies leads, personalizes outreach, books calls
Self-improving content system
$2K-5K/moAgent creates, tests, and optimizes content variations on a retainer
AI voice receptionist
$1.5K+Voice agent handles inbound calls, qualifies, books
Automated reporting
$2K-5KAgent pulls data, generates insights, ships weekly reports
GTM Labs runs a 12-agent open-source marketing pipeline (DevRel Origin) in production and installs it for teams as the DevRel Swarm Install ($4,500). For custom agent fleets on retainer, see AI Agents.
Ready when you are.
Discovery calls are 20 minutes. First one's on me.