Understand and improve.
Agents analyze selected workloads, modernize code in Rust and test the result against the original behavior. Every change reviewable.
THE LOOP
Observe. Prioritize. Change. Serve. Verify.
One loop, one data model. The evidence from the running system travels into the code change, the rollout and the next decision.
01 · OBSERVE
Assess
Billing, kernel telemetry and cloud configuration joined per function, task and instance.
02 · PRIORITIZE
Hotspots
Failure rate, latency, peak CPU and calls per day ranked by cost, inside the workbench.
03 · CHANGE
Modernize
Capture, plan, translate, check, repair, verify. A receipt per unit.
04 · SERVE
Deploy
Content-addressed kernels pinned to trust pools: shadow, canary, primary, rollback.
05 · VERIFY
Forecast & monitor
Outcome monitor, backtested forecast, evidence vault. The next decision starts with proof.
HOW IT THINKS
AI understands the code. Context informs the choice. Math evaluates the options.
Model what matters.
A ledgered fact graph connects systems, dependencies, ownership and operating rules, with history and traceable assumptions.
Evaluate where it fits.
Supported execution options compared on workload demand, hardware benchmarks, capacity, price and reliability. Trade-offs explicit.
01 · ASSESS
See beyond the bill.
Know not only how much. Know where. Know why. The Bunny Lab joins kernel-level telemetry with your billing APIs to reveal the cost, performance and security of every function, task and instance, in one view.
- Every dollar attributed down to the function, by service and region
- Services, dependencies and call paths discovered from real execution
- Idle capacity and security gaps, each with a number attached
IDE · WORKBENCH
One workbench, from the hotspot to the rollout.
The workbench is where the assessment turns into engineering. A native, GPU-rendered editor on a Rust foundation, with the hotspot list, the agent panel, the benchmark receipts and the deploy panel in the same window.
- Production hotspots ranked in the editor; one click hands a unit to an agent
- Claude, Codex, Copilot, Grok or any ACP agent behind one typed permission boundary
- Deploy panel: deploy to shadow, apply pin, roll back, with the approved receipt attached
02 · MODERNIZE
The rewrite, without the risk. Proven.
Every enterprise has code it can't afford to rewrite and can't afford to keep running. The engine rewrites it function by function into memory-safe Rust, in planned, reversible waves, and proves every unit against what production actually did.
- Golden inputs captured from production, not from a test plan
- Structure computed, not hallucinated: a deterministic plan the model only labels
- Compiler-as-oracle repair, differential verification, a ledger of every run and gate
03 · TARGET
A foundation you can hold accountable.
Performance problems surface under real load, exactly when they cost the most. Workloads land on a foundation where the same inputs produce the same outputs, every time, placed where they fit, with the record to prove it.
- Kernels are WASM components with one export and zero imports: nothing reaches the host that was not declared
- A ledgered fact graph holds what runs where, who owns it, and what constrains it, with history
- Placement compares supported options on cost per run; untrusted executors are verified by replication
AWSGCPAzureOCINutanixvSphereProxmox30D
Trust ledger
Executions4,812
Audits passed241 / 241
Divergences0
Idle TimeMap · node fleet-ba-112
free (harvestable)host-reserved (p95)
04 · FORECAST
Model what happens next.
Static, backward-looking plans miss the feedback loops that drive overruns and panic buying. By the time a shortfall shows in a report, the cheap decision window has closed.
- A forecast that ships with its own backtest: the method, the history, and how wrong it has been
- Scenarios that put the team's assumptions next to measured demand
- An outcome monitor with three verdicts: healthy, flagged, breach
05 · DEPLOY
Locked in to no one.
Chip architectures are multiplying, roughly five-fold by 2035, and every vendor's pricing depends on how hard it is for you to leave. The same workload runs anywhere and moves when the economics change, without touching the application.
- One workload on x86, ARM, RISC-V, GPU targets and WebAssembly at the edge
- Work grains: one content-addressed module over one content-addressed shard, on any machine you enroll
- Pins born in shadow, promoted up a guarded ladder, rolled back by the outcome monitor
$ bnl deploy --target fargate ✔ live 00:41 $ bnl deploy --target wasm-edge ✔ live 00:12 $ bnl deploy --target core ✔ live 00:23
PLATFORM
What every stage stands on.
IDENTITY & ENTITLEMENTS
RBAC, SSO, signed entitlement leases verifiable offline, and a service account for every enrolled node.
AGENTS & MCP
Typed tools behind one permission boundary: list_hotspots, first_golden_case, placement_quote, report_kernel_candidate, rebuild_grain, deploy_grain.
EVIDENCE
Append-only ledgers everywhere: facts, runs, costs, gates, executions. Checkpoints signed with RFC 3161 timestamps.
ANY SCREEN
Native desktop, browser, iOS and Android builds of the same workbench, and a console for the compute grid.