Market expectations
What enterprises expect from a gateway, and where Flexway stands.
Every expectation of today and tomorrow, each with Flexway's status in the published release; what is already there for a partial one, when unfolded.
Upcoming releases are not commitments; "Delivered" only when the catalogue of the published release says so.
Today 44 expectations
| What enterprises expect | Flexway v1.6.3 |
|---|---|
| Centralised, secure access to models | |
| Centralise access to providers and internal models | Delivered · since v0.1.0 |
| Hide provider keys, manage secrets | Delivered · since v0.1.0 |
| Authenticate applications, users and services (OIDC, JWT, mTLS) | Delivered · since v0.1.0 |
Rights per team, application, model or environmentAlready there, and what is missingScopes per endpoint, overrides per tenant and per agent, models allowed per backend; models allowed per caller (tenant, plan, use case, client) are missing | Partly · next in 1.9 |
Forbid direct access to providersAlready there, and what is missingOnly the gateway holds the keys; forbidding the direct network route is the job of the customer's network egress rules | Partly · next to be scheduled |
| Routing between models | |
Choose the model by taskAlready there, and what is missingRouting by capability ( | Partly · next in 1.9 |
| Fail over when a provider is down | Delivered · since v0.1.0 |
Choose by latency, cost, quality or regionAlready there, and what is missingCost: a soft cap that switches; measured latency, declared cost, region and hosting class as declared criteria are missing. Quality cannot be measured by the gateway (evaluation: out of scope) | Partly · next in 1.9 |
Send simple requests to a small modelAlready there, and what is missingPossible through a declared route (the task says so); a classifier on the path is excluded by a product rule | Partly · next in 1.9 |
Tell cloud, private and local models apartAlready there, and what is missingThe | Partly · next in 1.9 |
| Cost control | |
Quotas per application, team, user or projectAlready there, and what is missingTenant and agent built; user and use case framed; application, team, project and cost centre are missing | Partly · next in 1.8 |
Limit requests and tokensAlready there, and what is missingRequests per second per caller, tokens per day and per month; a token rate per minute against peaks is missing | Partly · next in 1.8 |
Monthly budgets and overrun alertsAlready there, and what is missingMonthly budget in euros per tenant, gauges and status in the console; budget alert thresholds and built-in rules are missing | Partly · next in 1.8 |
| Control peaks, prevent abuse | Delivered · since v0.1.0 |
Refuse, queue or reroute on overrunAlready there, and what is missing
| Partly · next in 1.8 |
Internal chargeback per cost centreAlready there, and what is missingConsumption per tenant, CSV export from the console; attribution per cost centre and its export are missing | Partly · next in 1.8 |
| Data protection with models | |
Detect or mask personal dataAlready there, and what is missingMasking of REST response fields built ( | Planned in 1.9 |
Prevent confidential data from reaching some providersAlready there, and what is missingA request marked sensitive is pinned to an internal backend; data classes and the backends allowed per class are missing | Partly · next in 1.9 |
Data residencyAlready there, and what is missingAnnounced "by policy", not built: regions allowed per tenant or per class, proven in the trace | Planned in 1.9 |
Filter prompts and responsesAlready there, and what is missingInput and output guardrails in the catalogue from the start, never framed | Planned in 1.9 |
Detect injection attemptsAlready there, and what is missingDeclared patterns and the customer's guard service; the gateway is not a sufficient guarantee (the analysis says so too) | Planned in 1.9 |
Trace filtering decisions and refusalsAlready there, and what is missingChained audit of authorisation decisions, closed refusal codes; the guard decision line, without content, is missing | Partly · next in 1.9 |
| AI observability and audit | |
| Latency, availability, errors, tokens, estimated costs | Delivered · since v0.1.0 |
Link each call to an application, user, use caseAlready there, and what is missingTenant, agent, client; hashed subject in the audit; use case framed; application and cost centre | Partly · next in 1.8 |
| Correlate traces with OpenTelemetry | Delivered · since v0.1.0 |
Analyse models used and error ratesAlready there, and what is missingPer backend and capability; the model served is in the audit, not in the metrics | Partly · next in 1.8 |
Audit evidence and compliance reportsAlready there, and what is missingChained audit; exportable execution proof planned; DORA and NIS2 reports on the horizon | Partly · next in 1.9 |
Follow cost and quality after a model changeAlready there, and what is missingCost per backend, reload marks, comparison of two windows; quality belongs to an evaluation platform | Partly · next to be scheduled |
| Agent tool governance | |
| Control agents' tool calls | Delivered · since v0.1.0 |
| Expose business APIs as MCP tools | Delivered · since v0.1.0 |
Authorise by identity and contextAlready there, and what is missingScopes, roles, unattended agents, delegation chain; contextual control of arguments (an amount by identity) is missing | Partly · next in 1.10 |
| Limit each agent's tools | Delivered · since v0.1.0 |
| Trace agent -> tool -> API -> system | Delivered · since v0.1.0 |
Require human approval for a sensitive actionAlready there, and what is missingThe gateway requires and verifies an approval proof issued by the identity provider; the workflow stays with the customer | Planned in 1.10 |
| Enterprise use cases | |
| Turn existing APIs into AI tools | Delivered · since v0.1.0 |
Control autonomous agents (look up, not pay)Already there, and what is missingScopes per endpoint, | Partly · next in 1.10 |
Govern RAG access (search, vectors, embeddings)Already there, and what is missingEmbeddings under budget, search APIs and vector databases served as endpoints; filtering rights per document is out of scope | Partly · next to be scheduled |
Industrialise the prompt lifecycleAlready there, and what is missingConfiguration versioned in Git, models allowed per backend; models approved per caller planned; prompt registry and result comparison: out of scope | Partly · next in 1.9 |
Cut latency and costs (cache, small model)Already there, and what is missingREST response cache; semantic cache planned; routes per task planned | Partly · next in 1.9 |
Cloud and self-hosted side by sideAlready there, and what is missingSensitive requests pinned to an internal model, relay to another Flexway; data classes and hosting planned | Partly · next in 1.9 |
Agent identities, limited rights, validity periodAlready there, and what is missingDeclared agent client, delegated token bounded in duration and depth, an exchange that only ever narrows; rights limited to one task are missing | Partly · next in 1.10 |
Govern AI in IDEs and CI/CDAlready there, and what is missingEvery client goes through the same identity, quota and audit; attribution per application (the client) with its budget is missing | Partly · next in 1.8 |
Catalogue of AI servicesAlready there, and what is missingCatalogue of endpoints, Agents screen (tools, federated servers), declared backends; a unified catalogue (models, agents, tools, owners, terms) is missing | Partly · next in 1.10 |
Regulatory rules by sensitivityAlready there, and what is missingPinning, audit without prompts; models, data and logging per class planned; execution proof planned | Partly · next in 1.9 |
Tomorrow, by 2030 14 expectations
| What they will expect | Flexway v1.6.3 |
|---|---|
| Emerging uses | |
Semantic routing, selection by task, sensitivity, budget, latencyAlready there, and what is missingPlanned through declared rules; autonomous selection by a model is excluded | Partly · next in 1.9 |
MCP gateway and agent governanceAlready there, and what is missingApproved tools = contract, scopes, quotas, inter-agent traces built; contextual arguments, budget per run, registry to come | Partly · next in 1.10 |
Dynamic authorisation, trusted delegationAlready there, and what is missingBounded RFC 8693 delegation built; task rights and approval proof planned, together with the IAM | Partly · next in 1.10 |
Automatic optimisation of cost, quality, latencyAlready there, and what is missingRecommendations as reviewed proposals (AI for operations); autonomous application is excluded: the console never mutates | Partly · next in 1.10 |
Model, prompt or agent versions: canary, rollbackAlready there, and what is missingVersions of an upstream; canary between inference backends; prompt versions: out of scope | Partly · next in 1.9 |
Federation of gateways across entitiesAlready there, and what is missingMulti-gateway console framed with no milestone; common policies distributed through GitOps; shared common fragments: candidate | Partly · next to be scheduled |
| Trajectory 2026-2030 | |
2026-2027 "industrialise": authentication, keys, token quotas, multiple models, costs, observability, first MCP usesAlready there, and what is missingFoundation built; cost attribution and declared routing planned | Partly · next in 1.8 |
2027-2028 "govern": tools, agent identities, contextual authorisation, delegation, audit, versionsAlready there, and what is missingTools, delegation, A2A built; agent actions planned | Partly · next in 1.9 |
2028-2030 "optimise": federated policies, dynamic selection, arbitration of cost, performance, riskAlready there, and what is missingFederation: candidate; arbitration by declared rules only | To be scheduled |
| Predictions | |
The gateway arbitrates agents' actionsAlready there, and what is missingThe contract's endpoint is the business action; approval and task rights planned | Partly · next in 1.10 |
Edge, cloud and device cooperation; central policies applied locallyAlready there, and what is missingThe binary runs locally with no central plane; signed configuration pulled by a site: planned | Planned in 1.11 |
Decisions tied to the real state of the systemAlready there, and what is missing
| To be scheduled |
Trusted intermediary between organisationsAlready there, and what is missingMultiple issuers, partner mTLS, token exchange, per-field encryption; declared purpose of use: candidate | Partly · next to be scheduled |
"Immune system": revoke, suspend, block a version, isolate, distribute a rule, protect locallyAlready there, and what is missingRevocation by reload, closed by default, egress guards, declared automatic rollback, local protection by design; an automatic suspension decided by a detector is excluded (proposals only) | Partly · next to be scheduled |
What Flexway does not do, and who it leaves it to
What the market analysis and Flexway place outside the gateway.
| Responsibility | Left to | What Flexway does at the border |
|---|---|---|
| Authenticate users and services | IAM, identity provider | Consumes and verifies tokens; never issues any |
| Agent logic and state | Agent runtime, orchestrator | Governs A2A without terminating it |
| Document search, per-document rights | RAG platform | Governs access to search and embedding APIs |
| Evaluate answer quality | LLM evaluation, LLMOps | Exposes costs, latencies, errors per backend for that evaluation |
| Prompts, models, versions | LLMOps registry and pipeline | Enforces the declared approved list and the canary decided elsewhere |
| Human approval of a sensitive operation | Business workflow | Requires and verifies the approval proof |
| Analyse traces and incidents | OpenTelemetry platform | Emits OTLP; the console covers what the gateway sees |
| Secrets and machine identities | Vault, IAM, PKI | Secret references, the customer's KMS, mTLS; SPIFFE on the horizon |
Let's talk about your gateway
Flexwiz is a young company: you talk directly with the person who founds it. Tell us what you want to run through Flexway. We talk about your infrastructure and your digital sovereignty requirements.
Online booking is coming soon; until then, an e-mail is enough.