Work architected at Fifth9
The platform
underneath the work.
Connect useful AI experiences to the infrastructure they depend on: identity, model access, work coordination, conversation interfaces and evidence. Each service has its own responsibility and maturity.
Scope: Service implementations and a dated development-environment review.
Status · 3 November 2026 review snapshotThe status descriptions below attribute observations to a separate 3 November platform review. They are not fresh runtime verification by this page. Services span internal and development environments, with different deployed versions and incomplete paths.
One front door does not mean one maturity
Shared sign-in.
Different capabilities behind it.
The review reported six public MCP endpoints sharing identity metadata. Their health responses did not establish successful tool use.
Detailed diagram and text
Reported development interface relationships
MCP client
A tool follows the advertised access contract.
Identity service
The endpoints name a shared sign-in service.
Memory
Five reported deployed tool definitions; authenticated calls not demonstrated.
Quality
Testing interfaces; predictive features need separate evaluation.
Templates
A read interface whose complete backend path was not established.
Agents, pipeline and team
Reported in-memory shells without mapped backends.
A map of advertised access relationships and reported limits, not an integrated execution trace or a production deployment claim.
Illustrative evaluation
A healthy endpoint.
Then the real task.
A tool discovers an endpoint and receives a healthy response. The next question is whether an authorised person can complete the intended operation and recover its result. A shell or an incomplete backend path must remain visible.
“Can someone finish the work—not just reach the service?”
Keep discovery, authentication, operation and outcome evidence separate.
Technical reading
The services and their boundaries.
Reported observations, implemented interfaces and complete user journeys carry different evidence.
How to read the dated status labels
- Running · reported snapshot
- The review reported observing the service doing work within its sampled scope.
- Deployed, limited · reported snapshot
- The review reported an available deployment with a material limitation. Full user journeys may remain unverified.
- Code only
- Source defines a capability; runtime availability was not established.
- Shell · reported snapshot
- An interface reported in-memory operation without a mapped backend; it is not evidence of a durable service.
These labels describe the review’s evidence, not a continuing availability commitment. Health, metadata and counters have narrower meaning than an independently exercised user journey.
Identity, models and work coordination
- Identity · reported running in development
- Published metadata describes signed-token issuance and agent-facing sign-in, alongside service credentials, token exchange, registration and revocation. The review did not establish all advertised flows through complete client journeys.
- Model gateway · reported running internally
- An entry point to model providers, with chat, streaming and embedding interfaces. In the inspected chat code, an omitted provider can default to a configured provider; a model mismatch can substitute that provider’s recommended model. Both are signalled. Strict configuration rejects those cases. Unknown providers, or mismatches without a recommended model, are rejected. Runtime configuration and useful cost accounting need separate verification.
- Work router · reported running internally
- Source and the review describe work placement, claims, heartbeats and recovery, with issue-tracker and message-bus intake. Optional model-assisted decomposition has its own usage and evaluation scope. Work placement is not proof of completed execution.
Conversation, catalogue and persona services
- Webchat · reported deployed with limits
- A conversation interface with streaming, history and persona selection. The review described a deployed version with unresolved alignment to the main code line and incomplete team-chat integration.
- API portal · reported deployed with limits
- A contract-based operations catalogue with tool mappings and citation-supported explanations. Its documented operations and internal availability are distinct from verified execution of every operation.
- Personas · reported running internally
- Defined assistant configurations, with review and rollback mechanisms. A persona is a configuration, not evidence of expertise or a trained personal model.
Certification receipts and result evidence
The review described a staged service-evaluation pipeline with signed verdict receipts. Its observed sample contained twenty blocked runs and fourteen errors, with no passing run in that sample. This is a dated observation, not proof that no certification has ever passed elsewhere.
Shared-secret HMAC receipts and Cosign artifact attestations are different mechanisms. The inspected signing code attempts Cosign only when enabled, the verdict passes, the executable is available and the image digest is known; success still depends on the signing result. A receipt’s existence does not establish a passing verdict, artifact correctness or production readiness.
The six reported MCP front doors
- Memory · /memory · deployed, limited
- Five tool definitions were reported in the deployed version, against seventeen in newer source. Read the memory interface case.
- Quality · /quality · deployed, limited
- The review described eight testing and quality tools. Failure prediction and test healing were not supported by benchmark evidence supplied for this page.
- Templates · /templates · deployed, limited
- The review described four read-only registry tools but an incomplete path to the backing service. A responding front door does not establish usable template retrieval.
- Agents, pipeline and team · shells
- The review reported /conduir-agents, /conduir-pipeline and /conduir-team as in-memory interfaces without mapped backends. Such shell state should not be treated as durable.
The development domain is dev.conduir.ai. This page reports the dated review; it has not repeated endpoint probes, authenticated tool calls or service-level evaluation. No service listed here is presented as a production product offer.
Continue exploring