Authority travels with the action. Evidence travels separately.
Policy is set centrally. Each interaction carries bounded authority and is verified on arrival at the boundary it reaches. Nothing returns to the middle.
THE ARCHITECTURAL INVERSION · MSEP
Distributed execution shouldn’t depend on central execution authority.
Most agent governance sends every consequential action through a gateway, or back to a central service for an answer. That puts one shared dependency in the critical path of every branch: fan-out becomes gateway load, and every join waits on the slowest round trip. MSEP, the protocol inside Mira, inverts it: bounded authority travels with each interaction, and the boundary it reaches verifies it locally.
Every action waits on a remote answer. If the service is slow or down, the work stops with it.
Every call funnels through one choke point, and that choke point holds the keys to every system behind it.
Authority is checked on arrival, beside the system being called. Nothing central sits in the path.
Centralise administration where it helps. Distribute execution authority where it matters.
Latency isn’t the whole argument. MSEP preserves the topology of the autonomous system itself, so governance follows the work as it fans out, delegates and reconverges, instead of forcing every branch back through the middle.
What that one change makes possible.
Parallel stays parallel
Branches are verified at their own boundaries at the same time, so governance doesn’t serialise the work. About 0.57 ms per governed hop in pre-production; over a ten-hop chain in equal-work tests, 1.8 to 2.1 times sooner than an inline gateway and 2.2 to 3.0 times sooner than a central call-out.
Keeps working when a service doesn’t
There is no central service for every action to wait on. Routine decisions are made where the action lands, and evidence is sent separately, so it never holds up the work.
Zero bypass
Governance means nothing if an agent can simply use another credential. Agents hold no standing keys to the systems they call: each boundary withholds the protected capability until the action is permitted.
The workflow is governed, not just the step
A permitted action doesn’t make the workflow it creates permissible. Mira governs who may take part, what must come before what and how a run may recover, and seals the graph a run actually took beside the one it was allowed.
Recover, don’t restart
Where recovery is safe and defined, Mira restores trusted state and continues, with an approved substitute agent, model or provider where policy permits, rather than restarting the whole run.
A Red Card follows the agent
An agent that forges or tampers with its authority is stopped and issued a Red Card, which signals actor-wide quarantine across your estate via SSF/CAEP. It constrains what that agent may do next until a named person reinstates it, so trust doesn’t reset when the workflow does.
Evidence you can verify and keep
Every decision at every boundary is sealed into verifiable causal evidence, sent apart from the action so it never slows it down. The payload stays in your own storage, and auditors can check the record without trusting Liora.
| Question | Gateway or central service | Mira boundary |
|---|---|---|
| Where authority is checked | In the middle, away from the action | At the system the action lands on |
| Across a multi-step workflow | Every step goes back through the middle | Authority travels with the work, bounded at each hop |
| Network traversals per action | 2 to 3 | 1 |
| Who holds system credentials | The gateway, for everything behind it | Each boundary, for its own system only |
| If the central service fails | Agent work stops | Routine decisions carry on |
| Decisions available | Allow or block | Six outcomes, from release to recover |
| Evidence | Logs at the checkpoint | A record anyone can verify, for every decision |
| Across parallel fan-out | Queues at the shared gateway; tail latency compounds | Verified concurrently at each branch boundary |
BUILT.
WORKING.
MEASURED.
WORKING.
MEASURED.
0.57 msp50, a complete governed hop plus successor state
83 µsp50, inbound authority verification
212 µsp50, evidence sealing
50+representative pre-production scenarios, from payments to pharmacovigilance
Pre-production measurements on our disclosed test environment, not production SLAs. Patent pending.
AI agents now approve payments, change production systems and move customer data. Most governance only finds out afterwards. Mira makes the decision before the action runs.
Identity isn't authority
Knowing which agent acted doesn't prove that this action, on this system, was allowed at that moment.
Central checkpoints slow everything
Sending every action to a remote service adds delay, fragility and lock-in as workflows grow.
Logs don't stop actions
An audit trail explains what happened afterwards. You also need to stop the unsafe action, and prove the control you applied.
Governance at every consequential boundary.
Mira makes the workflow governable; MSEP makes governance effective at each consequential execution boundary. This is what a boundary does each time an action arrives.
Checked before it runs
When an agent tries to pay, export, change or send something, the boundary checks it against your rules first. Then it allows it, removes private details, holds it for more authority, or stops it.
Private details stay private
Names, birth dates and contact details are removed before they reach a model or vendor that shouldn't see them.
Policy decides when more authority is needed
That authority can come from a named person, another trusted system or another source your policy defines. Nothing runs until it arrives, and a timer running out is never a yes.
Recovers where it is safe to
Where recovery is safe and defined, Mira restores trusted state and continues on an approved alternative model or provider, rather than restarting the whole run.
Fits the stack you run
Works with any agent framework, model or cloud, including agents you built yourself. Change a provider without rebuilding your controls.
Evidence your auditors can verify
Every run leaves a tamper-evident record of what was checked, decided and done, mapped to the regulations you answer to. Auditors can verify it for themselves, without having to trust us.
NATURAL-LANGUAGE RULEBOOK BUILDER
Write the rule in plain English. Mira makes it enforceable.
Risk, compliance and quality leads don’t need to write code. Describe the policy the way you would say it, link it to the regulation it comes from, and Mira compiles it into a deterministic boundary your agents cannot argue with.
- 1Write it in plain English
The way you would put it to a colleague.
- 2Map it to its legal basis
Each rule carries the regulation or policy clause it comes from.
- 3Compile it to a deterministic boundary
Explicit rules you can read, test and sign off.
- 4Test, see the impact, sign off
Nothing takes effect until an authorised person publishes it.
RULEBOOK · DATA HANDLING
YOU WRITE
Do not send customer health information to an external model. If the classification is uncertain, a privacy officer decides.
LEGAL BASIS
GDPR Art. 9 · health dataPrivacy Act 1988 · APP 6Data policy §4.2
COMPILED RULES
PII-1Health information to an external model
interdictPII-2Classification uncertain
elevatePrivacy officer✓Deterministic · versioned · signed offThe sentence stays with the rule. Only the rule is enforced.
Nothing improvises at runtime. A model can help you draft a rule. Only the compiled rule is ever enforced, the same way every time.
Plain English
For risk, compliance and quality leads. Describe the policy, review the rules it produces, and sign off.
Visual workflow canvas
For process owners. Draw the workflow, choose which agents act on what, and mark where a person decides.
Three lines of Python
Wrap the agents you already run and govern every step. See the code
All three build the same rulebook, so a compliance lead and an engineer are always looking at the same rules.
PLUGGABLE DECISION INTELLIGENCE
One reasons. One executes.
Mira doesn’t try to be your sanctions screen, credit model or clinical engine. Plug in the one you already trust. It reasons and signs its verdict. Mira acts on that verdict, for that exact action, and nothing else.
- No verdict, no action.
- A verdict meant for a different action, or one that doesn’t check out, is refused.
- The engine stays the expert. Mira stays the boundary.
Inside Mira.
Real screens from an example drug safety workspace, using synthetic data.
The same failure. With Mira, and without.
A credit-approval workflow where an agent leaks personal data mid-run. Switch Mira on to see what changes. One of 50+ scenarios in our demonstration build spanning banking, credit, pharma, healthcare, payments, software deployment, and more.
SCENARIO 03 · CREDIT APPROVAL · PERSONAL DATA LEAK
Without Mira, the leak is invisible. The workflow completes, the record looks clean, and nobody ever knows.
Mira in ninety seconds.
What Mira checks, what it stops, and the evidence it leaves. Plain English, captions on.
Built for the teams who answer for AI.
Security, risk and compliance leaders need controls they can explain to a board and evidence they can hand to an auditor.
Independent of your AI vendors
Works across models, clouds and agent frameworks. Change a provider without rebuilding your controls.
Your data stays yours
Customer data stays under your control. Liora keeps the proof, not a copy of your data.
Control before the action
Consequential actions are checked against your rules before they happen, not explained after the fact.
Evidence anyone can verify
Tamper-evident records that auditors can check for themselves, without having to trust Liora.
Fits the controls you have
Open interfaces for the identity, policy and security tools your organisation already trusts.
Keeps working under pressure
Routine decisions don't wait on a central service, so a slow or failed dependency doesn't stop the business.
DESIGNED WITH THESE IN MINDEU AI ActISO/IEC 42001NIST AI RMFGDPRAPRA CPS 230 / 234UK Corporate Governance Code (Provision 29)21 CFR Part 11 / EU Annex 11OpenTelemetry
Govern the agents you already have.
- Keep your stackInstrument existing agent workloads without replacing your models, frameworks or infrastructure.
- Debug with real evidenceSee what happened, in what order, and which control was applied.
- Apache 2.0 SDKOpen-source instrumentation for the agents you already run.
BUILD FOR FREE. GO LIVE WHEN YOU'RE READY.
Always-Free Developer Licence
Full access to the Mira engine, plain-English rulebook builder, workflow designer and Audit Vault for development, sandboxes and client PoCs.
Consultancies and engineering teams keep 100% of the rulebooks and workflows they build.
Mira engineRulebook builderWorkflow designerAudit Vault
A production licence applies when you go live. Full licence terms will be published with the developer release.
The 60-second developer surveyAnonymous. Tell us what's blocking your agents and we'll publish what we learn.
Take the surveyagents.py
# Three lines. Your agents, now on the record.from mira_agent import govern app = govern(app, policy="credit-approval")app.run() # every step governed
SIMPLIFIED EXAMPLE
From working build to production.
Mira is built and demonstrated. We are hardening it for production now, with design partners joining ahead of the developer release.
NOW
Architecture demonstrated
A working build showing governance, interdiction, recovery and verifiable evidence across agent workflows.
NEXT
Production hardening
Independent benchmarking, security review, enterprise deployment tooling and framework integrations.
THEN
Developer release
The Apache 2.0 SDK and the Always-Free Developer Licence: the Mira engine, plain-English rulebook builder, workflow designer and Audit Vault, free for development, sandboxes and client PoCs.
BEYOND
Enterprise governance
Policy management, assurance, reconstruction and sovereign deployment for consequential AI workloads.
Patent pendingOnboarding design partners
Put your agents on the record.
Join the design-partner programme. Early partners help shape Mira and get first access to the platform and SDK.
- Early access to the platform and SDK
- A direct line to the founding team
- A say in what we build first
Prefer email? hello@liora-ai.co
Request early access
Tell us a little about what you're building. We'll be in touch before the developer release.