01
Classify
Scores task type, complexity, context shape, and failure cost before every call.
Steroids for
your harness.
The drop-in optimization layer that cuts agent fleet model spend by up to 90% without touching your orchestration logic.
routing trace live
prompt"Refactor auth middleware for SAML SSO"
classifytask.class=refactor · quality.floor=0.90
route→ sonnet-4.6bypassed opus-4.1
cost$0.0042was $0.0191 — saved 78%
outcomesame quality · less spend
One line changes everything
+ client = OpenAI(base_url="https://api.hamblygmen.space/v1")Drop into any agent stack
LangChainClaude CodeCodexHermes AgentPi AgentCursorDevin
01 / Architecture
Between the harness and every model call.
One endpoint. No agent rewrite. Hamblygmen observes the signals your harness already produces and makes a better execution decision in milliseconds.
Agent harness
Your existing stack
LangChainCrewAIHyperlaneCustom
< 8ms overhead
Hamblygmen proxy
Agent-aware execution layer
01Real-time classification
02Dynamic quality-bar routing
03Semantic cache + deduplication
Model providers
Any model estate
OpenAIAnthropicDeepSeekLocal / vLLM
02 / Benchmarks
Provable, not promised.
The built-in eval runs 45 rubric-graded agent calls through four routing strategies. Same prompts, same gateway, same grading — only the routing differs.
45-item rubric-graded eval● sample run
| Strategy | Pass rate | Cost | vs frontier |
|---|---|---|---|
| frontier-only | 91.1% | $0.366 | baseline |
| random routing | 80.0% | $0.118 | −11 pts quality |
| cheapest-only | 53.3% | $0.002 | −38 pts quality |
| hamblygmen | 82.2% | $0.147 | 90% quality · 40% cost |
Run it yourself: node src/cli.ts eval — the eval is built into the gateway, not a marketing page.
02 / Runtime
Every call earns its model.
Optimize for the actual work—not a static provider preference or one-size-fits-all model tier.
02
Route
Selects the fastest, lowest-cost model that still clears the required quality bar.
03
Cache
Detects semantically equivalent work across agents, loops, and retries before spend occurs.
04
Learn
Builds a private performance map for each harness and task pattern over time.
03 / Fleet economics
Model the savings on your fleet.
Adjust the operating profile. Estimates use 400 runs per agent per month and 2.2K tokens per turn.
500 agents
18 turns
80%
Illustrative estimate. Actual performance depends on task distribution, prompts, providers, and quality constraints.
Current monthly spend
$34,294
With Hamblygmen
$6,516
81%
projected cost reduction
Latency gain
40%
Cache hit rate
18%
04 / Category
A layer, not another framework.
Your orchestration remains the source of truth. Hamblygmen makes each model call more efficient and dependable.
| Capability | Hamblygmen | Generic LLM router | Orchestration framework |
|---|---|---|---|
| Agent loop awareness | — | Varies | |
| Quality-bar routing per task | Basic | — | |
| Semantic cache + deduplication | Optional | — | |
| Replaces your agent architecture | — | — | |
| Learns per-harness performance | — | — |
05 / Deploy
Built for a quick test. Ready for the fleet.
Start with one client call, then take the same control plane into your most demanding environments.
From install to first optimized call in minutes.
One-line model client endpoint swap
Python and TypeScript SDK patterns
Local evaluation and shadow mode
Per-call routing traces and cost telemetry
policy.trace
// unchanged harness code
+ OpenAI(base_url="https://api.hamblygmen.space/v1")
→ classified
→ routed
→ observedOptimization compounds with every call
Stop sending every task to your most expensive model.
Keep the harness you built. Add an intelligence layer that learns where every request should run.