Control plane · v0.1 · agent-native
Evidence becomes
governed action.
Celest is the agent-native control plane for the Microsoft cloud. Discover every agent across Copilot Studio, Microsoft 365, GitHub, and Azure DevOps — then run one canonical lifecycle from evidence to approval to a verified receipt. The same loop, for the humans and agents operating the estate.
Every claim on this page is labeled by evidence. Live · read-only Experimental Roadmap
The rebirth
You knew Celest as
the Vercel of Flutter.
In YC W24, Celest let Flutter developers declare their backend in Dart and deploy it like magic — removing the seams between what a builder intends and what the cloud requires. The insight holds. The layer changed.
The builder is now a human-plus-agent organization. The cloud narrowed from generic infrastructure to the Microsoft substrate where enterprise identity, communication, work, code, data, and automation already live. Same founder. Same instinct. One layer up.
- Flutter developers forced across fragmented cloud toolchains. Operators and agents forced across fragmented Microsoft control planes.
- Cloud Widgets made infrastructure declarative. Canonical plans, authority, and receipts make operations inspectable.
- Generated clients connected app code to the cloud. API, MCP, CLI, and UI adapters connect agents to governed capability.
- Hid accidental complexity, never the developer's intent. Normalizes Microsoft complexity without erasing provider-native power.
The problem
One agent, scattered across
a dozen admin centers.
Microsoft is assembling an agent platform across Entra, Microsoft 365, Copilot Studio, Power Platform, Foundry, Agent 365, GitHub, and Azure DevOps. Each surface knows a sliver. None of them knows the agent.
- Identity
- Contoso Renewals Copilot ↳ Studio + Entra
- Owner
- unknown ↳ no maker record
- Lifecycle
- published · assigned · idle 41d
- Cost
- unknown ↳ credits unattributed
- Risk
- unknown ↳ oversharing not evaluated
- Source health
- 2 sources degraded
The highest-value gap is not another inventory screen. It is the transition between knowing and doing — today performed by scarce specialists, improvised in spreadsheets and tickets, or skipped. As agent count and autonomy rise, that becomes the bottleneck.
The lifecycle
From knowing to doing —
one governed loop.
The interface is replaceable. The durable product is the canonical lifecycle and the evidence it accumulates. Every consequential step has an inspectable actor, scope, decision, and policy basis — and closes with a receipt, not an API “success.”
- 01Observation
A change or scheduled sweep triggers the loop.
- 02Evidence snapshot
Immutable, with source health and provenance.
EvidenceSnapshot - 03Finding
A deterministic finding or an agent decision.
Finding - 04Proposal
Proposed actions and their predicted effects.
LifecyclePlan - 05Authority
Explicit approval or a policy-based decision. The gate.
AuthorityDecision - 06Execution
Provider-native mutation within bounded scope.
- 07Receipt
An immutable record of what actually happened.
ExecutionReceipt - 08Reconcile
Fresh observation closes the finding — or reopens it.
Evidence before action.
Unknown is not zero.
Authority is data.
Receipts close the loop.
What's true today
We hold ourselves to the same
evidence standard we sell.
Unknown is not zero. Here is exactly what is built, what is experimental, and what is still roadmap — no invented traction. The current MVP is a read-only Dev deployment, not a multi-tenant production service.
The map
The Microsoft AI
Operations Stack.
Five layers of operating outcome around one canonical estate. We lead with AgentOps — the registry, lifecycle, authority, and receipt layer — and expand outward from there.
Start here
The Agent Estate Review.
A two-week, fixed-scope engagement. Read-only. Software gathers and normalizes the evidence; together we turn fragmented registry, cost, security, and ownership signals into an operating plan — then become the recurring AgentOps layer underneath it.
Inputs
- A bounded tenant / environment scope
- Supported read access to relevant Microsoft sources
- Interviews with platform, security & business owners
- Optional licensing, credit, GitHub & Azure DevOps evidence
Outputs
- Normalized agent inventory with source & join provenance
- Ownership, lifecycle, tool, MCP, usage, cost & risk findings
- A posture report that distinguishes zero, false, unknown & unavailable
- Prioritized work with evidence and expected outcomes
- An authority & operating-model workshop
- A 30/60/90-day AgentOps plan
Then it expands
- Recurring managed Estate Review
- LicenseOps & CreditOps reporting
- GuardOps & identity remediation
- Governed work-item creation
- Lifecycle execution after explicit control gates
Why this founder · why now
Built the platform.
Now Customer Zero.
Dillon created Amplify Flutter at AWS, watched developers struggle with the seams between application and cloud, and built the original Celest around that pain — through YC W24. He now operates Microsoft as both a builder and an administrator, running a real tenant.
Celest is Customer Zero for the agent-native operating model it sells: agent identity and sponsorship, permission tiers, findings, approvals, interventions, and receipts — captured as evidence. The product came from direct reverse-engineering and use, not a market map.
Why the window is open
- Microsoft is building the substrate — but the customer still owns the seams between products.
- Agent 365 is positioned as an observability, governance, and security trust layer.
- Copilot Credits make agent usage an operating-cost concern, not just a seat concern.
- GitHub ships agentic workflows; Azure DevOps holds a huge installed base with no equivalent.
Microsoft will improve its native dashboards. Celest wins on cross-surface workflow, evidence continuity, customer policy, provider breadth, and speed — not on a temporary private endpoint.
The loop is open.
Celest is taking a small number of design partners for the first paid Estate Reviews. If you run a Microsoft-centric estate — or serve one as a partner — and agent sprawl is becoming your problem, let's talk.
Read-only, fixed-scope, paid pilot. No data leaves your control without an explicit gate.