Reasoning
What should I do?
- GPT
- Claude
- Gemini
- Llama
- Qwen
- Mistral
NewThe state layer for AI
Your AI knows what happened. Does it know what’s true now?
AI agents now work across hours and thousands of steps. As tasks evolve, facts change, instructions are revised, and decisions get superseded. Cricket continuously verifies and reconciles what’s happening, maintaining the current state your existing models need to keep working.
More tasks completed. Less wasted compute. Same models.
How your AI works today
- Input
- Model
- Tools
- Output✓
CricketLive state
- 01ObserveCapture what just happened
- 02VerifyCheck it against evidence
- 03ReconcileUpdate what’s true now
- 04DeliverCurrent state to the output
Event
Customer: “Let’s move it to Monday.”Check
Deadline → Monday✓Customer emailMonday✓Project ticketMonday✓ SupportedCurrent state4 items
- DeadlineFridayCurrent
- ↳Launch planShips FridayCurrent
- Budget40kHeld
- ApproverTeam ACurrent
SupersededBudget: 35k
The team behind Cricket, as featured in
Two ways to use it.
The missing layer in AI
AI is getting smarter. Longer tasks need more than intelligence.
Models can now take on complex work. But as that work runs for hours, days or weeks, facts change, instructions evolve, and every decision depends on what came before.
Memory helps AI remember. State tells it what still applies.
Memory
What happened before?
- RAG
- Vector DBs
- Long context
- Mem0
- Zep
- Summaries
State
What’s true right now?
Cricket
Cricket provides the state layer for your existing models and infrastructure.
What happens as AI tasks get more complex?
As changes and dependencies pile up, AI without a dedicated state layer becomes less reliable.
Successful task completion
100%80%60%40%20%0%The gap Cricket closesExisting AI system + CricketExisting AI system
Existing AI + CricketExisting AI system
- Simple tasksLimited state
- Multiple dependenciesMore tools, files and decisions
- Changing requirementsConflicting or updated information
- Long-running workContinuous updates and decisions
Increasing changes and dependencies within a task
More tokens in context, so quality naturally degradesMore tokens in context, lower quality
What state unlocks
Get more out of the models you already use.
Cricket adds verified state to your existing AI stack, helping models complete complex work more reliably and efficiently.
- 01 / EfficiencyUp to 22×Lower cost per successful taskGet more finished work out of the same AI budget.Cost per successful taskTodayWith state22× lower
- 02 / ReliabilityUp to 6.7×Stronger policy adherenceMore consistently follows instructions and constraints across long-running tasks.Policy compliance rateExisting AI systemBaselineSame system + CricketUp to 6.7×
- 03 / Accuracy+69 ptsHigher accuracy beyond the context windowThe longer the task gets, the more normal systems degrade. Cricket holds up.Accuracy82% with Cricket13% withoutTask length / context load →
- 04 / Model leverage~10×Smaller models, generations aheadCricket layered on smaller models performs multiple generations ahead at a fraction of the cost.Smaller model + Cricket91% correct$0.06Frontier model alone100% correct$0.65Competitive performance. Far lower cost.
Works with your stack
Keep everything you’ve built. Cricket layers on top.
The industry is pouring billions into models, fine-tunes, memory and infrastructure. Swap any of it, as often as you like. Cricket stays the same underneath, keeping every layer working from what’s true right now.
Agent frameworkLangGraph
ModelGPT
MemoryMem0
Serving & cachevLLM
CricketState layerAlways on
observe → reconcile → serve
Output
Correct next action ✓
- ✓Current facts
- ✓Rules held
- ↓Cost per task
Already have fine-tunes, memory or RAG? Good. Cricket makes them work harder.
Fine-tunes teach a model how to act. Memory and RAG recall what happened. Neither keeps track of what’s true right now. That’s the layer Cricket adds, on top of whatever you run.
- 01Your stack, unchanged
Your model, fine-tune, memory and tools run exactly as they do today. Nothing to rip out or retrain.
→ - 02Cricket watches
An observer reads every event alongside your calls: messages, tool results, edits and policy changes.
→ - 03Cricket works out what’s true
Memory recalls what happened. Cricket decides what still applies: what changed, what was superseded and what’s still open, with evidence.
→ - 04Everything you have gets sharper
Compact current state goes into every next call. The same model, fine-tune and memory, now working from what’s true.
Minimal added latencyNo retrainingNo migration
See it happen
Same technology underneath. What it looks like in your product.
As instructions change, Cricket keeps your agents on the current state, so they avoid costly mistakes and rework.
+Your workflowlegal, support, research, sales, ops…
AI agents & automation
Same task. Less wasted work.
As instructions and dependencies change, Cricket keeps your existing agents aligned with the current state, helping them avoid costly mistakes and unnecessary rework.
Shared taskBuild and publish the Harbor Coffee website: menu, checkout, free delivery over $25.
Both reach the same result. One needed a rollback, a second pass and a human to get there.
Swipe to follow both paths →
Existing agentWithout Cricket
Rework loop
Task assigned
Homepage, menu, cart
Rule: $25 → $40
Checkout built with old $25
Checkout test fails
Moves on to styling
Publishes
Customer reports errors+ human time
Roll back+ tool calls
Reopen finished files+ tokens
Fix tax and banner+ tokens
Re-test+ compute
Republish+ execution
Live
Same agent + CricketCurrent state maintained
Task assigned
Homepage, menu, cart
$25 superseded → $40 current
Checkout built with $40
Test failure tracked · publish blocked
Styling
Fix tax
Re-test passes
Publishes once
Live
Without CricketPublished after rollback and rework
With CricketPublished correctly on the first attempt
Tokens129k105kincl. Cricket overhead
Tool calls2418
Elapsed time61 min23 min
Human interventions10
Illustrative values, not a measured run. Your evaluation measures the real difference, including Cricket’s own overhead.
Same finish line. Less work to get there.
Avoiding state-related mistakes means fewer retries, less wasted compute and a lower cost per successfully completed task.
Cricket doesn’t make your model remember more. It helps your model know what still applies.
FAQ
The questions technical teams ask first.
Why can’t I just have my agent keep a state file or a summary?
You can, and for simple workflows that may be enough. It gets hard as tasks evolve, sources conflict and several systems contribute updates: something has to decide what supersedes what, what’s uncertain, and what each step actually needs. Cricket handles that maintenance and reconciliation as dedicated infrastructure, so you don’t rebuild it for every workflow.
What happens when Cricket gets the state wrong?
Every update is reconciled against evidence, conflicts are flagged as uncertain rather than silently overwritten, and every value carries its history and source. You can inspect what Cricket believes, see why, and correct it, and corrections flow to every agent using that state.
How is this different from memory, RAG or a longer context window?
Memory and RAG preserve and retrieve what happened; a longer window lets a model see more of it. None of them decide what is still true. Cricket maintains the current state of the task, including what changed and what was superseded, and works alongside your memory and retrieval rather than replacing them.
How much latency and compute does Cricket add?
Cricket runs alongside the agent loop and serves a compact state, so the overhead per step is small relative to re-processing history. Every evaluation reports the latency overhead and total cost, including Cricket’s own, next to task success.
What data does Cricket store, and where?
Only the task state and the events you choose to send. Data is isolated per customer, with configurable retention, and Cricket can run in your cloud or ours.
Do I need to change my model or retrain anything?
No. No weights change. Cricket works with the models, tools and frameworks you already use, and you can swap models at any time.
Give your AI a sense of now.
Bring one long-running workflow. We’ll run it with and without Cricket and show you the difference in task success and cost per successful task.