Why ThetaLayer
One plane for every agent, vendor and system of record
Your agents run on different platforms, act under different identities and write to different systems. ThetaLayer reconciles all of it into one record per agent: what it did, what it cost, and whether it was worth it.
- Agent runtimepost_invoice → successCan't see what the ERP actually recorded
- ERP change logposted by SVC_AP_AGENTCan't see which run did it, or who asked
- Approvalsno request on fileCan't see that one was required
- Model provider bill€592.56 · OctoberCan't see what any of it produced
The problem
Every system holds part of the answer. None holds all of it.
Ask who did what, at what cost, with what result — and you get four partial answers from four teams. That's why most organizations still can't say what their agents are worth.
already limit AI use because of token and running costs
McKinsey, The state of AI in 2026
have a mature governance model for autonomous agents
Deloitte, State of AI in the Enterprise 2026
- 01
Agents act across systems
One invoice run reads an email, appends a tracker row, checks the vendor in the ERP, posts the invoice and requests an approval. That's five systems, five logs and five different identifiers for one piece of work.
- 02
Systems of record don't know an agent was involved
The ERP records a posting by a technical user — or by the clerk whose delegated credentials the agent was using. It can't tell you which run made the change, or who asked for it.
- 03
Runtimes don't know what actually changed
An agent's log records what the agent believes happened. A timeout may have posted the invoice anyway; a reported success may never have landed. Nobody finds out until someone reconciles by hand.
- 04
Cost bills don't map to outcomes
Token invoices, platform credits and seat licenses arrive monthly, per vendor. None of them says what a completed invoice costs once failures, retries and sub-agents are counted.
One plane · four pillars
A mandate, a ledger and a scorecard for every agent
The same four records for every agent, whichever vendor built it and whichever systems it touches.
01 · Mandate
What each agent may do
Tools, data, amount thresholds, spend caps, autonomy level and identity rules — written down per agent and versioned.
- A draft mandate per agent, checked against its full history
- Approval thresholds, sensitive-field rules and spend caps
- Autonomy levels from suggest-only to autonomous
“Invoices over €10,000 need an approved request before posting.”
02 · Ledger
What each agent actually did
Every run reconstructed end to end and reconciled against your systems of record — with what it cost and what it touched.
- Each claimed action matched to independent evidence
- Agent, person and credential identified separately
- Metered and allocated cost per run, sub-agents included
“Agent reported a timeout; the ERP shows the invoice posted — then the retry posted it again.”
03 · Scorecard
What it produced
Cost per completed workflow, straight-through and takeover rates, evidence coverage and mandate compliance — comparable across vendors and versions.
- One set of metric definitions for every vendor
- Regressions tied to the version that caused them
- Routing decisions by task type, not by vendor
“€0.70 → €2.40 per completed invoice after v2.4 shipped.”
04 · Lifecycle
RoadmapWho owns it
A named owner, an autonomy level and a quarterly review for every agent, backed by its scorecard and exception history.
- Named owners and review dates
- Autonomy raised or lowered on track record
- Renew, re-scope or retire as an explicit decision
“Quarterly review: renew, re-scope or retire.”
Examples are from a demo environment with simulated data.
Who it's for
Built for the people who own the outcome
Agent decisions cut across budget, operations and control. ThetaLayer gives each owner the same evidence, framed for the decision they have to make.
- Every agent, on every platform, in one inventory with cost per outcome
- Evidence to scale, renew or retire each agent at review
- Build-vs-buy decisions on measured cost, not vendor claims
“Which of our agents earn their cost?”
- Agent spend allocated to teams, workflows and versions
- Cost per completed workflow, including failures and retries
- Approval and payment controls evidenced, not asserted
“What do we pay per invoice processed — and is it rising?”
- An exception queue with the evidence attached
- Straight-through and takeover rates by task type
- Each task type routed to the agent that handles it best
“Where are people still cleaning up after agents?”
- The agent, person and credential behind every change
- Mandate compliance re-checked whenever a rule changes
- An audit trail of every human review decision
“Did an agent change something it shouldn't have — and who approved it?”
How it compares
Different tools answer different questions
Vendor control planes govern agents inside their own ecosystems. Security tools protect them. APM debugs them. ThetaLayer reconciles what every agent did and cost against your systems of record — alongside all of them.
| Capability | Accountability planeThetaLayer | Vendor control planes | AI security tools | Observability & APM | Spreadsheets & BI |
|---|---|---|---|---|---|
| Works across agent vendors and runtimes | Yes | PartialInside their ecosystem | Yes | PartialInstrumented apps | PartialManual |
| Reconciles agent claims with systems of record | YesConfidence per link | PartialTheir own records | No | No | PartialManual sampling |
| Cost per completed business outcome | YesIncl. failures, sub-agents | PartialOwn consumption | No | PartialTokens per trace | PartialMonthly, manual |
| Mandate checks, re-run across history | Yes | PartialPolicies in-platform | PartialRuntime guardrails | No | No |
| Two-sided evidence for audit | Yes | PartialIn-platform logs | PartialSecurity findings | PartialTechnical traces | No |
| Runtime protection (prompt injection, data leakage) | No | Partial | Yes | No | No |
| Real-time enforcement inside the agent platform | PartialLater, at chosen tool boundaries | Yes | PartialGuardrails | No | No |
| Latency and error debugging | No | Partial | No | Yes | No |
Based on the typical scope of each category; individual products vary. Most enterprises will run several of these side by side — ThetaLayer reads from them rather than replacing them.
What ThetaLayer reads from and sits alongsidePrinciples
How we build it
Neutral
We don't build agents or sell models. Every vendor is measured with the same definitions, so comparisons between them hold up.
Evidence over assertions
An agent's log is a claim. Every metric traces back to records in your own systems, with the link method and confidence shown.
Read-only first
Start by observing. Add enforcement later, only at the tool boundaries you choose, once the evidence supports it.
Privacy by design
Sensitive values are hashed or dropped inside your network before anything is sent. Credentials never leave it.
Book a Demo
See your agents through one plane.
Book a 30-minute walkthrough. We'll reconstruct a real workflow end to end — mandates, evidence and cost per outcome — and scope a six-week, read-only pilot around your agents.