Skip to content

Product

A control layer, not another endpoint.

Waymesh sits between your application and the providers that execute its work. It decides where each workload should run, enforces the policy it must run under, and keeps the record that shows whether the decision was right.

  1. Intelligent routing
  2. Provider abstraction
  3. Policy controls
  4. Observability
  5. Evidence-driven deployment

Capability

Intelligent routing

Every request is scored against the routes that could serve it, on predicted cost, observed latency, measured reliability and expected quality for that kind of work.

The weighting is per workload, not global. A five-thousand-call extraction job and a multi-step reasoning task are not the same question, and a single set of weights would answer both badly.

Candidate scoring

workload: long-document summarization
  • Anthropic$0.04128.4 s0.894
  • OpenAI$0.05087.1 s0.881
  • Private / self-hosted$0.011122.6 s0.798
  • Groq$0.00631.2 sExcluded

Illustrative example — not customer production results.

Capability

Provider abstraction

Applications submit a workload and its requirements. Provider credentials, model identifiers, endpoint selection and per-provider quirks stay inside the control layer.

That is what makes a provider change a configuration change. Retries, fallbacks and token accounting stop accumulating in your application as conditionals nobody wants to own.

POST /api/agent/jobs

{"taskType": "TEXT_ANALYSIS","privacyTier": "PRIVATE","priorityMode": "BALANCED","maxBudget": 0.25,"input": "…"}

No provider. No model name. No endpoint. The request states what the work needs; the control layer decides where it runs.

Capability

Policy controls

Privacy tier, budget ceiling and capability requirements are evaluated before scoring, as pass-or-fail gates.

A constraint is not a heavily weighted preference. A workload that may not leave your infrastructure has no path to a hosted API, however well that route would otherwise have scored.

Policy evaluation

evaluated before scoring
Privacy tier
Private — execution on infrastructure you control
Budget ceiling
$0.25 per request
Capability
Context window ≥ 48,200 tokens
Routes considered
4
Routes eligible
1
Routes gated out
3 — before any score was computed

Illustrative example — not customer production results.

Capability

Observability

Routing produces a record: spend and traffic by provider, latency and reliability as observed rather than advertised, and the decision behind every executed job.

Predicted cost is stored next to what was actually settled, so an estimate can be checked rather than trusted.

Provider operations

last 30 days
Groq
$46.00 · 41,820 requests · 420 ms
OpenAI
$175.22 · 18,640 requests · 2.2 s
Anthropic
$200.57 · 12,305 requests · 3.9 s
Private / self-hosted
$4.49 · 6,410 requests · 9.7 s
Total
$426.28 · 79,175 requests

Illustrative example — not customer production results.

Capability

Evidence-driven deployment

Waymesh can run in shadow mode: it evaluates every request and records the route it would have chosen, while your existing routing continues to control execution.

Authority follows evidence. Routing takes control of real traffic only once the record supports it, and only for the segments that record actually covers.

How shadow and authoritative operation differ

Path to authority

  1. ObserveShadow deployment
  2. CompareShadow deployment
  3. ValidateAhead
  4. PromoteAhead

Promotion is gated on evidence and is not enabled by default.

Category

Passing a request along, and deciding where it goes.

Both behaviours are useful and they are not the same job. A gateway gives you one place to call. A control layer gives you a decision, made per request, that you can inspect afterwards. Many teams want both, and the distinction is worth being precise about when choosing.

Who picks the model
Gateway The caller names it in the request
Control layer The layer selects it from what the workload requires
What travels with a request
Gateway A provider and model identifier
Control layer Requirements: task type, privacy tier, budget ceiling
When pricing matters
Gateway Reconciled after the fact, on the invoice
Control layer Costed per candidate before dispatch, settled against it after
What a policy does
Gateway Enforced by the caller, if at all
Control layer Evaluated as a gate before any route is scored
What is recorded
Gateway The call that was made
Control layer The candidates considered, the rejections, and the reason for the winner
Failure handling
Gateway Retry or fall back in application code
Control layer Eligibility is re-derived per request from live route facts

Waymesh is the second column. It is not a drop-in replacement for a gateway you already run — it decides which route a workload should take, and can sit in front of whatever executes it.

The most useful way to evaluate a routing decision is against your own traffic, in shadow mode, where it changes nothing.

Apply for Design Partner Access