InstaControl™ · Agent control plane

Accuracy gets an agent approved. Evidence gets it deployed.

Who authorized this, what did it touch, what did it cost, and can you prove it six months from now. InstaControl™ answers all four, for every action, before it runs and after.

Where it sits

Three layers. You control the top one.

Knowledge underneath, execution in the middle, your controls on top. The same three layers whether the work is quoting, dispatch, or field service.

Policy is configuration, not a release. Your team changes what agents may do without filing a ticket, and without waiting for us.

  1. Control and auditInstaControl™Who may act, what it costs, what it did.You control this
  2. ExecutionInstaWorkers™The agents and workflows that do the work.
  3. Knowledge and memoryInstaBrain™What the business knows, kept current.
Every action an agent takes passes through the top layer, including agents you built yourself.
One plane, two jobs

Deciding and proving are the same surface.

Most teams buy these separately and find the gap in the incident review: a policy tool that cannot see runtime, and a dashboard that can see everything and stop nothing.

Control before the action

Evidence during and after

  1. Policyruntime, not prompts
  2. Accessroles and purpose
  3. Approvalroute and timeout
  4. Budgetcaps per capability
  5. Reversibilityrollback handler
  6. erp.updateQuoteThe actionallowed, held, or declined
  7. Tracehash-chained steps
  8. Costtokens and latency
  9. Evalson your datasets
  10. Outcomerevenue and SLA
  11. Audityour Postgres
One plane. The rule that stops an action and the record proving it was stopped are the same object.
Instrumented, not estimated

What one request writes to your database.

Agent spend is not a monthly invoice surprise. The gateway wraps the model call itself, so cost and latency belong to the request, the session, the user, and the workflow that caused them.

  • Ships as / GatewayA Python package that wraps your LLM calls and writes telemetry to a Postgres you already run.
  • Ships as / InfrastructureA Terraform stack for dashboards, alerting and retention. Or skip it and use your own.
agent_eventsone request2.63s
  • Context412ms
  • Model1020ms · ttft 240ms
  • Tool call1200ms
  • Costinput_tokensoutput_tokensusd_costsession_cost
  • Latencytime_to_first_tokentime_to_last_tokentotal_duration
  • Configmodel_idtemperaturetop_ptop_kmax_retries
  • Identityprojectenvironmentuser_idsession_idevent_id
  • Toolstool_namearguments_classruntime_ms
  • Payloadrequest_sizeresponse_size

OpenAI · Anthropic · Google, one schema, your database

How it goes live

It runs on your traffic before it runs on your customers.

Promotion is a decision someone on your side makes, against numbers, on the record.

share of live traffic
0%10%50%100%a regression rolls back
  1. ShadowReal traffic copied in. Nothing exposed.
  2. CanaryFirst live outputs. Acceptance rate starts counting.
  3. ExpansionHeld against accuracy, latency, cost, review rate.
  4. StandingFull traffic, still gated on every release.
Architecture review

The questions your team asks in the first hour.

Single-tenant, in your environment, with isolation enforced at the storage layer rather than by an application filter someone can forget.

Is our data training a shared model?
No pooling, no cross-customer training. Your data trains your models only, and the knowledge stays yours if you leave.
Does this replace our observability stack?
No. It captures the layer your APM cannot see, and exports downstream to the tools you already page on.
Does it work with agents we built ourselves?
Yes. The gateway wraps model calls from your own code, and the plane covers any action registered against it.
What happens when an agent gets it wrong?
The action was reversible by contract, the approver who let it through is named, and the trace shows what it reasoned over.
Where this is going

The plane is the part that does not get replaced.

The model, the provider, and the machine underneath will each change several times inside the life of one deployment.

The control layer has to survive that, or you re-approve your AI program every time the stack moves.

Stays put

  • Who may act
  • What needs sign-off
  • What it cost
  • The trace
  • The eval gate
  • The audit record
  • Frontier model
  • Provider and price
  • Cloud, on-prem, edge
  • Large or distilled
  • Which workers
  • Workflow definitions

Expected to change

Routing sends each task to the model and the environment best suited to it. Your policies are configuration on top of that, not code inside it.
  • Shipping

    Authoring moves to your side

    Policy editing, approval design, and a no-code builder your operators can read without an engineer in the room.

  • Next

    A control worker, not just a console

    Agents that audit, scan for compliance drift, and validate policy continuously, instead of a quarterly review that samples.

  • Next

    The same plane at the edge

    An air-gapped rack gets the same policy surface and the same audit record as your VPC, with no raw data leaving the site.

  • Next

    Replay, not reconstruct

    Step back through a live episode exactly as it ran, and approve from voice or mobile when it cannot wait.

Everything on the plane

What your team gets to touch, and when.

A filled dot is running in deployments now. Ask us where a roadmap item stands before you build a plan around it.

12 live12 next

Live todayNext

  • Cost, latency, attribution5 of 8 live
    • Per-request cost and tokens
    • TTFT, TTLT, total duration
    • Every tool call and its runtime
    • Model config per call
    • Cost by project, env, user, session
    • Spend caps
    • Approved-model catalog
    • Outcome attribution
  • Traces, records, memory1 of 4 live
    • Telemetry in your Postgres, SIEM export
    • Trace viewer with hash chain
    • Memory and knowledge controls
    • Episode replay
  • Policy and access0 of 3 live
    • Nothing yet
    • Policy editor
    • Role and attribute access, purpose binding
    • No-code workflow builder
  • Approvals and reversibility1 of 3 live
    • Approvals inside deployed workflows
    • Approval design: predicate, route, timeout
    • Checkpoints and rewind
  • Evals and release3 of 4 live
    • Continuous evals on your datasets
    • Release gates that roll back
    • Shadow traffic and canary
    • Retirement console
  • Deployment and ops2 of 2 live
    • Single-tenant cloud, on-prem, air-gapped
    • Prometheus, Grafana, Alertmanager
    • Complete
Common questions

The short answers.

1Is this LLM observability with a new label?

Observability tells you what happened. It cannot decline an action. InstaControl™ evaluates policy before the tool call, holds the action for an approver when a threshold is crossed, and then records both facts in the same trace. The dashboard is one half of the product.

2We already have SOC 2 controls and change management. Why another plane?

Because your existing controls assume software behaves the same way twice. Agent behavior emerges at runtime, so the source of truth moves from the code to the trace. This is the plane that makes runtime behavior reviewable with the process discipline you already run.

3How much of this is on us to implement?

Instrumentation is a package install and a database connection. The policies, thresholds, and approval routes are decisions only your team can make, and we facilitate them during deployment rather than shipping defaults you would have to discover.

4Can we run it without egress?

Yes. The observability stack is built on components that run in your environment, and the design excludes anything that cannot operate air-gapped at the edge.

5Who can see a trace?

Whoever your access rules say. Traces are subject to the same role, attribute, and purpose rules as the data inside them, and views are derived from the underlying chain rather than assembled by hand.

6You will change models and providers. What breaks on our side?

Nothing you authored. Routing decides which model and which environment suits a task; your policies, thresholds, and approval routes sit above that as configuration rather than code. The evals are how you verify that a swap, including a swap to a smaller and cheaper model, did not quietly cost you accuracy.

7What do we own at the end of the contract?

The audit history, the policies, the evaluation datasets, and the knowledge your corrections built. It is in your database, in your environment.

Behind this solution: Lily, the AI Forward Deployed Engineer, working on the platform Lily builds on, delivered through forward deployed engineering. InstaControl™ sits over what the platform executes and what InstaBrain™ knows, including the workflows behind InstaProspect and InstaQuote, and it runs on the Small Data Center where the edge requires it.

Put an agent live that you can defend in a review.

Bring the questions your security and architecture teams will ask. We will walk the trace of a real action, end to end.