The runtime for the AI-native company
Your company runs on recurring operations — outreach, onboarding, monitoring, the back office. Omnislash puts AI agents in charge of them and holds them to it: durable through a crash, governed at every gate, verified against a definition of done, improving on their own.
You write the brain. The engine runs the operation.
Not a pitch deck — a runtime already running production operations. Planner, durable execution, governance, self-improvement, and the control plane: built and run end to end by one engineer.
runtime · execution loop
00 observe ->01 plan ->02 execute ->03 verify ->04 improve
! verify fails -> patch plan -> resume
What you can put on autopilot
Most company work is recurring operations. Put agents in charge of them — and each needs the same thing underneath: a runtime that's durable through a crash, governed at every gate, and verified against a definition of done. That's the layer teams hand-build and never get to hold. Omnislash ships it.
Pull knowledge out of email, Slack, tickets and DBs, keep it current, and turn it into something agents can act on.
Monitor against targets, decide, and adjust — without a human in the inner loop.
Outreach, support triage, compliance ops, back-office — a recurring service owned start to finish, humans on the gates.
Machine-readable interfaces — APIs, MCPs, CLIs — agents discover and use under permission.
01 · 02 · 03 · 04 → These aren't four products. They're one missing layer — a durable, governed runtime. Omnislash is it.
Brain vs Engine
Like a Next.js app to Vercel: your code, their runtime. Your logic stays yours — diffable, reviewable, self-hosted.
Brain — your git repo
Your domain. Your code. Versioned in your repo.
Engine — Omnislash
The runtime. Same for every brain.
The core loop
Not a slide. The engine's actual execution model — five real subsystems wired into one closed loop. It doesn't stop at "did it run?"; it checks "did it meet the goal?" — and when the answer is no, it edits the plan and goes again.
verify → improve · That last edge is the whole difference. The eval/observability field stops at a report a human reads. Here the failed check closes the loop itself — the plan is patched, the run resumes, completed work preserved. See how failures become edits →
Crash-proof by construction
Every agent is a durable object — one owner per agent, its state journaled, not parked in memory. Crash, redeploy, lose the network, and queued messages, HITL pauses, scheduled turns and stop requests all survive. Deterministic reconcilers sweep up orphans. Durability is the default, not a bolt-on.
durable run
run ▶ step 3 / 7
✕ crash (server killed, redeploy)
↻ reconcile orphan
run ▶ resume @ step 3 / 7 ← not step 1
mailbox · HITL pause · scheduled turn · stop request — all survive
Humans govern, agents work
Questions, approvals and capability requests all flow through one HITL gateway — one table, one event pair, first-responder CAS. Only the real exceptions reach you.
Self-correcting, not self-reporting
When a Definition-of-Done fails, a recovery agent diagnoses it, edits the plan blueprint itself, and resumes — completed work preserved. Failures become edits, not post-mortems. And completed runs become improvement PRs: validated by champion/challenger, auto-reverted on regression, and fenced against reward-hacking. It gets better on its own — and can't game its own score.
completed work preserved (01, 02) · the plan node is rewritten, not the run restarted
Any model, any harness — no lock
~200 models across 24 providers, plus CLI harnesses and OpenRouter. The chain keeps the work moving when a provider goes down.
fallback chain
model A ✕ (provider down) -> model B — same trace, not a restart
circuit: open · routed around across bot · worker · MCP
The proof is the running surface
Every screenshot here is the actual dashboard, running. Sandboxed by default (bubblewrap + OverlayFS), record-and-replay any run at zero token cost, steer any live agent mid-run and hand control back.
You never outgrow it
Graduate the same operation without a migration: AI end-to-end -> blueprint lock -> code-defined TypeScript pipelines . Same repo, same governance, durable state intact.
All of it — planner, durable execution, governance, self-improvement, this control plane — designed, shipped, and operated in production by one engineer.
Start with one operation
Pick one recurring operation that should run itself — durable, governed, verified. I'll map it to a brain and a scenario and walk you through how it runs on Omnislash. I build these, and I built the runtime they run on.
Prefer to dig in first? See the packages & image
Send me the one operation: what it is, who owns it today, and where it breaks. It comes straight to me — I read every one and reply myself.
Talk to me about your operationhi@omnislash.ai