Every Squidgy agent is an executable skills file: structured knowledge that an AI agent can act on, not just retrieve. Here's how that's built — and why it's a different category from RAG.
Every agent on Squidgy is composed from the same primitive — five files that together describe what the agent does, how it does it, and what it knows. Versioned like code. Inspectable like a contract. Runnable like a process.
YAML config
skills/{agent}/config.yaml
Capabilities, tools the agent can call (calendar, CRM, email, custom APIs, MCP servers), channels it deploys to, eval settings, deployment policy. The contract.
Example
name: maya
role: accountability_coach
tools:
- google_calendar
- hubspot
- slack
deploy_to: [embed, slack]
eval_suite: maya_v3
human_approval:
- send_to_customerSystem prompt
skills/{agent}/system-prompt.md
Voice, rules, escalation paths, behavioural guardrails. Written in plain English by Ace from your description, refined as the agent learns.
Example
You are Maya, an accountability coach for fitness professionals…
NEVER send a customer message without human approval.
ESCALATE to founder@example.com if a client mentions cancellation.
Voice: warm, concise, action-oriented. Use the customer's first name.Knowledge base
skills/{agent}/kb/
Your facts, policies, products, examples — your training corpus and your operating reality. RAG-grounded so answers come with citations to source documents.
Example
kb/policies.md # cancellation policy, refund policy
kb/programs.md # 8-week / 12-week / lifetime program details
kb/voice-guide.md # 12 example messages in your tone
kb/escalation.md # when to hand to a human, who to hand toWorkflow
n8n/{agent}_workflow.json
The orchestration — handoffs to other agents, branching logic, human-approval gates, retry policies. Runs on N8N under the hood; visible to you, executable by the agent.
Example
Trigger: customer message →
Pia router → routes by topic →
Maya draft reply →
Human approval queue →
Send via Slack/email →
Log outcome → evalDeployed agent
live across selected channels
Live across whichever channels you chose — embed widget, Slack, Teams, Roam, voice. Eval suites monitor for regression. Marketplace listing optional. Your skill files keep running as updates ship.
Example
Channels: web embed, Slack DM, Slack channel
Eval pass rate: 94% (target: ≥90%)
Marketplace listing: Certified · 4.8★ · 47 active customers
Last KB update: 2 days ago (cancellation policy)Both approaches use large language models. Both ground answers in your data. But they answer different questions. RAG helps a model retrieve. Executable skill files let an agent act.
| Dimension | Search over docs (RAG) | Executable skill files (Squidgy) |
|---|---|---|
| Primary job | Retrieve relevant snippets and compose an answer | Act on the customer's behalf — book, send, qualify, decide |
| Knowledge format | Documents indexed for semantic search | YAML + system prompt + KB + workflow — structured for execution |
| Inspectable? | Black-box composed answer | Every component diffable; agent state recorded |
| Versioning | Document-level (re-index) | File-level (skill files versioned like code) |
| Multi-step work | Single Q→A turn | Workflows with handoffs, branches, human-approval gates |
| Multi-agent teams | One bot per docs corpus | Router (Pia) dispatches to specialists, agents hand off |
| Tool/action capability | Limited — usually "answer only" | Native — agent calls tools to actually do the work |
| Learning loop | Add a doc to the index | Eval suite catches regressions; experiments graduate to playbooks |
| When you outgrow it | Customer asks for an action — you build a separate workflow | Agent already does the action; you ship the next one |
| Defensibility | Anyone can re-build with the same docs | Your skill files encode HOW you operate — that's the moat |
RAG and executable skill files aren't competitors — Squidgy uses RAG inside its knowledge base. The difference is the layer above: RAG-only systems retrieve and stop; Squidgy agents retrieve, decide, act, and continue.
Real workflows need specialisation, handoffs, and shared state. On Squidgy that's built in — the orchestration layer is part of the platform, not something you wire up.
Pia router
A router agent receives every inbound request and dispatches it to the specialist best fit for the job. Your specialist agents stay focused — Pia handles the air-traffic-control.
Agent handoffs
When one agent needs to pass to another ("this is a billing question, route to Marcus"), the handoff carries the conversation state. The next agent picks up cold-start free.
State management
Per-user persistent memory + per-conversation context. Agents remember commitments across sessions. Privacy is enforced — no cross-user leakage by design.
Skill files aren't static. The same versioning that makes them inspectable makes them improvable — by you, by your team, by the platform. The loop has three steps.
01
Audits
Eval suites run on every change. Production traces are sampled, scored, and flagged when behaviour drifts. Customer feedback (thumbs, escalations, refund signals) feeds back into the audit.
02
Experiments
When the audit surfaces a problem or opportunity, experiments are scoped: a new prompt variant, a tweaked workflow, an updated KB. Each experiment has a hypothesis, a metric, and a result.
03
Playbooks
When an experiment wins twice, it graduates to a playbook — codified into the skill file or template, reusable across agents. The next builder benefits without rediscovering the lesson.
The same executable skill files run wherever your customers are — without rebuilding per channel. Channel adapters handle the surface; your skill file is the source of truth.
A company brain is the structured, machine-executable record of how a business actually operates — its skills, workflows, knowledge, and decisions — packaged so AI agents can act on it directly. It's distinct from a knowledge base or RAG system: a knowledge base helps humans (or chatbots) search documents; a company brain gives agents the operating instructions to do the work themselves.
RAG retrieves snippets from your documents and feeds them to a model that composes an answer. Squidgy agents are built from executable skill files — versioned YAML configs, system prompts, knowledge bases, and workflows that the agent runs. RAG answers questions about your company. Squidgy agents do work on behalf of your company. Different jobs.
On Squidgy, agents are coordinated by an orchestrator (we call ours Pia, a router agent). Agents hand off conversations to each other, share state, and follow shared workflows. Knowledge bases are scoped per agent or shared across the team. Every handoff is recorded so the next agent picks up where the last one left off — no asking the customer to repeat themselves.
It's a versioned bundle of everything an agent needs to do its job: a YAML config (capabilities, tools, integrations), a system prompt (voice, rules, escalation), a knowledge base (your facts, policies, examples), and a workflow (the multi-step orchestration). Together they're executable — the agent runs them, doesn't just retrieve from them. Inspectable, diffable, exportable, and never a black box.
Two reasons. First: your company brain is the part competitors can't copy — the specific way YOU run YOUR niche. Second: structured-knowledge agents compound. Every iteration improves the skill files; every new agent benefits from the team's shared workflows. RAG over docs gives you a search tool. Executable skill files give you an operating system.
Yes — explicitly, by policy. Skill files (YAML + system prompt), knowledge bases, workflows, and customer data are all exportable on demand. Squidgy is a platform, not a lock-in.
No. The platform handles all the structure for you — you describe the agent in plain English to Ace and it generates the underlying skill files. The structure exists so engineers, ops teams, and the platform can inspect, diff, and improve agents over time. You see plain-English summaries; engineers (or future-you) can read the underlying YAML if they want.
Apply for the Agent Builder early-access cohort. Hands-on onboarding included.