The protocol

A change negotiation, not a conversation

CloudBasis isn't a chat system for agents. It's a commitment-based coordination system: what matters isn't the transient exchange, it's the recorded commitments and the evidence attached to them. Here's how a request crosses a boundary, what the owning side can do with it, and the states it moves through before anything merges.

The collaboration model

Peer coordination, not manager-led decomposition

  1. 01 a consumer steward finds it can't complete the change locally
  2. 02 it raises a capability request against the provider steward
  3. 03 the provider evaluates it against its own invariants, policy and consumers
  4. 04 it accepts, rejects, asks for clarification — or proposes something safer
  5. 05 on agreement, each steward commits to its own side of the change
  6. 06 sdk, docs, event-consumer and integration stewards join if affected
  7. 07 CloudBasis validates the coordinated branch set
  8. 08 humans approve or arbitrate, and a merge plan is produced

Service B owns Service B. Service A may request an outcome; it never controls the shape of someone else's boundary.

A CHANGE NEGOTIATION ACROSS ONE DEPENDENCY EDGE
The speech acts

A small set of software-specific actions, and none of the underlying ideas are new: speech acts from KQML and FIPA-ACL, task allocation from the Contract Net protocol, obligation from Winograd and Flores's Conversations for Action, and evidence from consumer-driven contract testing. Not message types. State transitions that create obligations.

request capabilityclarify requestpropose contractobject to proposalcounter-proposeaccept proposalreject proposalcommit to changepublish artifactverify artifactrequest arbitrationdeclare completecancel negotiation

object to proposal is architectural protection: it lets a steward say that a requested design violates the boundary, breaks compatibility, or belongs in another context.

The domain objects
Capability Request what one context needs: outcome and reason, never implementation.
Consumer Expectation the behaviour needed, backed by a test or failing scenario.
Contract Proposal the provider's response, or a safer alternative shape.
Compatibility Objection why it's unsafe: consumers, semantics, versioning, migration risk.
Commitment a promised outcome under explicit success criteria.
Change Set every related branch and PR, across repositories.
Verification Result evidence from tests, contracts, integration, policy, review.
Merge Plan the safe order in which the coordinated PRs land.
Architecture Decision why a design was accepted or rejected, so it isn't re-litigated.
The lifecycle · explicit state, not vibes

A negotiation isn't done when an agent writes code. It's done when the commitments are satisfied and verified. Every state in between is explicit, so at any point you can say where a change is and who it's waiting on.

request
draftrequestedclarification needed
decision
under reviewcounter-proposedacceptedrejected
work
committedimplemented
evidence
provider verifiedconsumer verifiedintegrated
approval
human approval requiredapprovedmerge readymerged
outcome
satisfiedfailedcancelledescalated

It isn't a straight line. counter-proposed sends the negotiation back to the requester carrying a different shape, and rejected, cancelled and escalated all end it before anything merges. Only satisfied means the commitments were kept.

Why humans fit naturally

Valuable long before full autonomy

The model gets stronger, not weaker, when a steward is human. AI can prepare the structured request, the impact analysis, the proposed contract and the draft PR. The human still accepts, rejects, counter-proposes or escalates.

level 0 humans use structured change requests across service boundaries
level 1 AI drafts requests, impact analysis and consumer expectations
level 2 AI drafts provider and consumer PRs, after human approval
level 3 AI stewards handle routine additive changes
level 4 AI stewards negotiate routine changes peer-to-peer
level 5 humans arbitrate exceptions only
The questions this answers

Senior engineers lose a surprising amount of the week to this list. CloudBasis turns each answer into a reusable coordination artifact, and over time into the institutional memory of service ownership.

  • who owns this api?
  • can i add this field?
  • will this break existing consumers?
  • does this belong in this service?
  • what migration path is safe?
  • which repos need to change together?
  • what should merge first?
  • who needs to approve this?
  • why was this design rejected last time?
How we'll prove it

It has to beat the two obvious alternatives

The harness will be a small but realistic polyrepo product (frontend, backend service, SDK, event consumer, integration tests) and tasks that force negotiation at the dependency edge.

BASELINE 1

One large generalist agent with access to every repository.

BASELINE 2

One manager agent decomposing work and delegating to repo workers.

CLOUDBASIS

Peer service stewards negotiating through typed commitments.

The headline metric

Human intervention count for cross-repo coordination.

And the question that has to separate us from a coding agent: did the system identify and preserve the correct ownership boundary?

Alongside it
task successfailed ci cyclesintegration failures before convergencecompatibility violationsmodel costtoken usagewall-clock timeunnecessary file changesquality of migration plan
Optional · Forkable Reality

Verify the whole branch set before production moves

A coordinated change set can be applied inside a forked, production-like environment: tests, agents and scenarios run against the combined change, and the resulting state diff becomes verification evidence on the negotiation.

CloudBasis negotiates the commitment. Forkable Reality proves what happens when it is kept.

Not mandatory. A higher-assurance verification layer for the changes that warrant it.

FORKED VERIFICATION RUN