The Agentic OS

The operating layer for agentic AI.

xFlo is the platform where composable skills are built, governed agents are deployed, and organisational context compounds over time. Not a chatbot. Not a framework. The infrastructure layer everything agentic runs on.

Get started Talk to us

xFlo dashboard — projects, flows, and skills catalog

The platform

Build, Manage & Optimise Agent Operations Right Across Your Organisation

Every capability on xFlo is built, governed, and observed the same way. Walk through each, in the order they compose.

01 · Orchestration

Agents orchestrate. Skills execute.

Agents are not monolithic programmes. They orchestrate governed, typed, reusable skills from a shared catalog.

An agent decides which skills to invoke, in what order, with what inputs. Change behaviour by reassigning skills, not by rewriting logic. The platform compounds with every skill you build.

Workspace DirectorThree-tier delegation. Workspace orchestrators route to project agents, which delegate to skills. Scope ownership that matches how organisations actually work.
Background and scheduled executionAgents spawn enrichment tasks, run cron conversations, and pause on failure. Your team focuses elsewhere while agents stay productive.
One shared catalogImprove a skill once, every agent that uses it improves with it. Versioned, attributed, audited.

Orchestrates

Content MarketerCheckpoint

Research

Draft

Review

Publish

4 Skills from catalog

02 · Composable skills

Fully composable skills

Skills are the atomic unit of capability. Four levels add up. No level removes what came before.

Level 1 is plain instructions. Level 2 adds tools. Level 3 adds typed input and output contracts. Level 4 drops to custom code when the problem calls for it. Every level carries the same governance, cost attribution, and composability primitives.

Build once, use everywhereA skill built for one agent is immediately available to every other agent in the catalog.
Governed at every levelEven Level 4 custom code runs through trust gates, cost attribution, and version control.

Skill

Instructions

Tools

Typed I/O

Custom Code

03 · Builders

Two builders. Two ways to build.

Skill Builder configures individual skills. Agent Builder assembles agents from the catalog. Both paths produce production-grade, governed output.

Go manual when you want direct control of schema and routing. Go conversational when you want the Builder Agent to draft structure, suggest tools, and ask clarifying questions. Approve, edit, or reject each proposal.

Conversational skill creationThe Builder Agent can propose skills mid-conversation with inline HITL approval, turning real work into reusable capability.
Interactive proposal cardsEvery proposed skill, tool, or constraint appears as a reviewable card. Nothing ships without your approval.

Skill BuilderManual

Name

draft_release_notes

Level

Tools

LinearGitHubSlack+ Add

TrustSupervised

Agent BuilderConversational

Build me a content marketer who can draft, review, and publish blog posts.

Proposing 4 skills + trust tier. Review each:

ResearchProposed

DraftProposed

ReviewApproved

PublishApproved

04 · Cascading ontology

Configure once. Inherited everywhere.

The six-layer cascade is how xFlo distributes context across your organisation.

Account, Workspace, Project, Thread, User, Skill. Set a constraint, brand voice, or governance rule at the right level. Everything beneath inherits automatically. Override at any level; the cascade picks the most specific value at invocation.

Version control at every layerEvery instruction change is versioned with full diff view and one-click rollback. Six layers, six histories.
Live propagationCascade changes apply at the next invocation. No rebuild, no restart, no redeploy.

Cascades down

01Account

Global rules

02Workspace

Brand voice

03Project

Campaign goals

04Thread

Session context

05User

Personal prefs

06Skill

Per-skill tune

05 · Memory

Memory compounds with every interaction.

The Base Reality Agent builds and maintains organisational context as a continuous lifecycle.

Tenant memory persists facts, entities, and relationships that matter to the workspace. Procedural memory captures what worked and what failed. Episodic memory holds conversations and user preferences. Every skill execution reads across all three at invocation.

The 50th agent is smarter than the firstMemory compounds. Agents do not start from zero, they start from what the organisation has learned.
Zero PII leak architectureTenant isolation enforced at the database level through row-level security policies. Multi-tenant safe from day one.

Base Reality Agent

Observes, writes, retrieves

Tenant memoryFacts, entities, relationships

Procedural memoryWhat worked, what failed

Episodic memoryConversations, preferences

EarlierCompounds with every interactionNow

06 · Governance

Governance is structural. Not optional.

HITL gates placeable anywhere. Trust escalation that scales with demonstrated performance.

Agents earn autonomy. Supervised: every output reviewed. Checkpoint: autonomous with human review at defined intervals. Autonomous: operates within boundaries, escalates on ambiguity. You define the thresholds. The platform enforces them.

HITL on irreversible actionsEmail send, calendar create, financial transactions — every irreversible action gates through a human approval. No silent executions.
Immutable event storeEvery agent decision, skill execution, and human approval is logged with full context. Complete replay capability, always.

Earned autonomy

Supervised

HITL on every output

Checkpoint

HITL at intervals

Autonomous

HITL on exception

Oversight decreases · quality-gated

07 · Observe & scale

Full visibility. At every level.

Per-skill cost attribution at account, workspace, project, and skill level. Quality scoring per execution.

No black-box billing. If a project is running over budget, you see exactly which agents and skills are driving costs. If a team is underusing agents, you see that too. Quality scoring feeds configurable thresholds and auto-approval policies.

Compute-class budgetsLIGHT 30s, STANDARD 120s, HEAVY 300s, PREMIUM 600s. Every skill declares its class. No runaway jobs.
Selective-retrieval cost architectureLarge skill outputs compressed to refs. Artifact store separate from chat history. Surgical revision via edit_artifact.

Billing · this monthUSD

Account$1,247.32

Workspace · Marketing$681.40

Project · Q4 Campaign$420.18

Skill · Draft$180.24

Skill · Review$95.10

Skill · Publish$144.84

08 · Connect

Everything your organisation uses. Connected.

850+ toolkits via managed OAuth. Custom MCP servers for proprietary systems.

Gmail, Slack, Salesforce, GitHub, HubSpot, Notion, Linear, Microsoft 365, and hundreds more. No auth configuration required. Every connection governed, audited, and scoped to exactly what the agent needs.

Per-user OAuth in shared threadsSame project, same thread, different user credentials. Full audit trail tying every tool call to the right user.
Generic CMS publishingPayloadCMS, Ghost, and custom systems — every CMS talks the same MCP interface. Zero custom code per target.

850+ toolkits

Gmail

Slack

GitHub

HubSpot

Salesforce

Notion

Linear

Google Drive

Coming soon

On the roadmap.

Forthcoming releases coming soon.

In development

Code execution sandbox

Agents clone repos, run scripts, analyse data, and build spreadsheets and charts inside isolated containers. No API limits, no rigid tool surfaces, full computational freedom with a complete audit trail.

In development

xFlo Desktop App

Native macOS and Windows client with system tray, global shortcuts, and local file access. Route to local models via Ollama for offline, privacy-first agent work.

In development

Skill Lifecycle Agent

Agents that actively improve other agents. Learns from execution traces, diagnoses failures, proposes evidence-based fixes, and manages versions with before/after validation.