SnapFlow
Swarm Intelligence
Your autonomous workforce — the agent fleet, their skills, the standing rules you give them, and the shared knowledge base they plan from.
Overview
Swarm Intelligence (route /swarm) is the roster and rulebook for SnapFlow's autonomous workforce: your fleet of specialist agents, the reusable skills/runbooks they can call, the standing rules and guardrails you set for them, and the shared knowledge base ("Company Brain") they draw on when planning. It has four tabs: Fleet, Skills, Governance, Knowledge.
Fleet
Every primary swarm agent, its live status, and its recent activity — click a card to drill in. The agents are fixed and shared across the whole platform:
| Agent | What it does |
|---|---|
| 🏗 Terraform — Infrastructure Provisioner | Plans and applies infrastructure changes with guardrails and drift awareness. |
| 🔄 Pipeline — CI/CD Engineer | Builds, repairs, and hardens delivery pipelines for safer releases. |
| ⚓ Harbor — Container Registry Manager | Manages image publishing, scanning, and registry hygiene across environments. |
| 🛡 Sentinel — Security Auditor | Finds risk, validates controls, and recommends safe remediations. |
| 💰 Ledger — Cloud Cost Analyst | Tracks spend, flags waste, and recommends practical cost reductions. |
| 👁 Watchman — Observability Agent | Listens for incidents, diagnoses service health, and verifies recovery. |
| 🔐 Vault — Secrets Manager | Rotates secrets, checks secret exposure, and prepares secure credential changes. |
| 🗄 Migra — Database Migration Agent | Prepares schema changes, rollback paths, and safe database recovery steps. |
| 🚀 Deploy — Release Manager | Ships releases, coordinates rollback paths, and restores stable versions when needed. |
| 🚨 Responder — Incident Commander | Executes recovery playbooks, coordinates mitigation, and closes the incident loop. |
| ↩️ RollbackWatch — Rollback Supervisor | Monitors deployments and infrastructure changes for rollback conditions, and automatically reverts state when verification fails. |
Status chips use the same color language as everywhere else: Running Active Error Queued Idle (the default when there's nothing to do).
Clicking an agent card opens an inspector with a description, a natural-language dispatch bar — type what you want that specific agent to do and click Go (it shows "Dispatching..." then "Running" while busy) — a Recent Tasks feed of its latest runs, and, if one of its tasks needs a decision, an inline approval panel right there.
Skills
Reusable runbooks the swarm follows. When a plan step matches a skill's agent and trigger words, that skill's instructions are injected into the step and it still runs through all the normal guardrails — the Pre-Flight Manifest, approvals, the sandbox. Built-in skills ship with SnapFlow; you can add your own to encode your team's specific workflows.
Two sub-tabs: Built-in (N) and Custom (N). Skills are grouped by category — 🏗️ IaC, 🔄 CI/CD, 🛡️ Security, 👁️ Observability, 💰 Cost, 🔐 Secrets, 🗄️ Database, 🚀 Deployment, 🚨 Incident Response, 📋 Compliance, ⚡ Custom — and each is a collapsible row showing its name (tagged "BUILT-IN" where relevant), a one-line description, and its category. Expanded, a skill shows which agents use it (or "Any agent"), its trigger keywords (or "Always (when agent matches)"), and either its runbook text or, for MCP-type skills, the specific tool/server it prefers.
Adding a custom skill
On the Custom sub-tab, click + Add Skill:
| Field | What it means |
|---|---|
| Type | "Runbook" (a written procedure) or "MCP tool" (calls an external tool server). |
| Skill Name | Free text. |
| Category | One of: IaC, CI/CD, Security, Observability, Cost, Secrets, Database, Deployment, Incident Response, Compliance, Custom. |
| Description | Free text. |
| Agents | Toggle-chip multi-select from the Fleet roster — leave blank to allow any agent. |
| Triggers | Comma-separated keywords — leave blank to match whenever the agent matches, regardless of wording. |
| Runbook (Runbook type) | A markdown textarea injected into the matching step. |
| MCP Server / MCP Tool (MCP type) | Which tool server and tool to call. |
Click Save Skill ("Saving…" while busy) or Cancel. Custom skills can be removed later from their expanded row with Remove Skill.
Governance
Two related but distinct rule types, switched with an inner toggle: Directives (standing instructions for how the swarm should think and plan) and Policies (hard guardrails that block or gate specific actions).
Directives
Standing rules injected into every agent's plan before any tool call or task breakdown — write your team's operating policy once and SnapFlow enforces it on every run.
- Type a rule into the text box, e.g. "Always create a rollback plan before production changes." A live hint under the box grades its length: too short (<20 characters, "be specific"), good (≤150), getting long (≤220), or too long ("LLMs may deprioritize") — with a running character count out of 200.
- Click Add. New directives are active by default.
- Or expand "Quick-start templates" for eight one-click presets — rollback plans, destructive-ops caution, IAM changes, DB migrations, post-deploy verification, no secrets in code, cost thresholds, Terraform review — each already-added preset shows a checkmark and disables.
Each directive in your list has an on/off toggle, click-to-edit text, a priority number with up/down arrows (lower number wins if two directives conflict), an "Active in planning context" / "Excluded from planning context" status line, and a delete button. Directive changes take effect within about 5 minutes (they're cached).
Policies
Guardrails your team authors, enforced both when a mission is planned and again right before each command runs.
| Effect | What it does |
|---|---|
| Block | Stops a matching mission at the gate entirely. |
| Require approval | Inserts a human approval gate on a matching step. |
If any preset policies are still unused, a "Starter policies" card lists them with one-click Enable buttons. To write your own, fill in the New rule form:
| Field | What it means |
|---|---|
| Effect | Block or Require approval. |
| Environments | Comma-separated environment names this applies to — blank means any environment. |
| Resource keywords | Matches if any keyword appears in the target resource. |
| Action keywords | Matches if any keyword appears in the action being taken. |
| Message shown when it fires | The text shown to you when this rule triggers. |
Click Add rule. Each active rule shows as a card with its message, effect chip, a meta grid (Envs / Resources / Actions / Window), and Disable/Enable plus Delete buttons.
Knowledge — Company Brain
Your agency's private, accumulating memory. Everything here — what you teach it manually and what it absorbs from completed work — is ranked and injected into every agent's planning context.
A vitals row shows four numbers: Knowledge (entries, with how many are verified), Success rate, Experiences (episodic memory count), and Value generated (dollars and hours saved).
Teaching the brain something
Click "+ Teach the brain something…" to expand a form: choose Write (type it directly) or From URL (point at a source), fill in a Title and the content or URL, add comma-separated Tags, then click Add to brain. It's injected into every future agent run from that point on.
If SnapFlow notices a lesson worth keeping from recent successful work, it appears under Suggested learnings with a one-click "+ Save to brain" button.
The knowledge base
Searchable list of everything the brain knows. Each entry shows a status badge (Verified Draft Deprecated), whether it was 🤖 learned automatically or ✋ taught by you, its content (expandable if long), tags, a usage count, and action buttons — ✓ Verify, Deprecate/Restore, Edit, Delete. A collapsible Failure insights section (when failures exist) breaks down what's gone wrong recently by category.
Settings → AI Configuration has on/off switches for whether the brain is consulted during planning (KB retrieval, episodic memory, etc). This Knowledge tab is where you actually add and manage what's in the brain. See Settings → AI Configuration.