Aegisora runtime reference
How Aegisora governs autonomous AI agents: the decision pipeline, canonical authority, execution continuity, evidence and enterprise controls.
- 12
- pipeline stages
- 3
- decisions: ALLOW, BLOCK, ESCALATE
- 9
- aligned packages (v2.0)
- 115/115
- runtime tests (v4.0.0 notes)
Overview
Aegisora is a runtime governance layer that evaluates and enforces agent actions before they cross protected execution boundaries.
Securing the model is not enough once agents can call tools, providers and external systems. Aegisora secures what the agent is allowed to do: every action is evaluated, decided and enforced at the execution boundary, and every outcome is recorded as evidence.
A blocked or escalated action never reaches the provider or the tool.
When state, identity or authority no longer matches what was authorized, execution stops.
Allowed, blocked and escalated actions all produce correlated, auditable records.
Workspace isolation, RBAC and server-side authorization are part of the runtime.
Where Aegisora sits
| Layer | Role |
|---|---|
| Aegisora Runtime | Security, governance, policy, risk, decisions, enforcement. |
| Agent framework | Orchestration and workflow execution (for example LangChain and LangGraph). |
| Model provider | Model inference and generation. |
| Tools and APIs | External execution surfaces governed by Aegisora. |
Aegisora works with the AI stack developers already use rather than replacing it. It sits between the agent and the execution surfaces it can reach.
Governance pipeline
Every governed action follows one canonical chain, from request to audit. Select a stage to see what it guarantees.
CONTINUITY SEAL
The authorized action is bound to the security state it was approved under.
- Binds workspace, agent, execution identity, action, resource, tool, execution input and authority state.
- The seal has an immutable identity and cryptographic integrity validation; inputs are canonically normalized.
- A seal cannot be created or executed while a containment condition is active.
Continuity Seal and Effect Reconciliation extend the chain from deciding whether an action is allowed to proving the executed action is still the authorized one.
Decisions and enforcement
The security boundary is explicit: ALLOW, BLOCK or ESCALATE.
| Decision | Effect at the execution boundary |
|---|---|
| ALLOW | Execution proceeds and carries canonical correlation into the governed execution. |
| BLOCK | Provider and tool execution is prevented. The prevented, not-attempted state is preserved. |
| ESCALATE | Execution stops pending approval. Provider and model resolution do not occur. |
Enforcement state and execution outcome
Aegisora separates what governance decided from what actually happened. This lets you answer both questions for any action.
| Dimension | Values |
|---|---|
| Enforcement | not_executed, executed, prevented, escalated |
| Execution | not_attempted, succeeded, failed |
Canonical correlation identifiers
| Identifier | Purpose |
|---|---|
traceId | Ties one request across governance, enforcement, execution, evidence and audit. |
decisionId | Identifies each governance decision independently. |
executionId | Identifies the governed execution. |
evidenceId | Identifies the evidence record. |
correlationId | Runtime-generated correlation across the security record. |
sealId | Identifies the Continuity Seal (v4.0) and links it to evidence. |
Attacker-controlled correlation identifiers are rejected as canonical identity. Forged runtime identities fail closed.
Continuity Seal
Authorization answers whether an action is allowed. Continuity answers whether it is still the same action, under the same authorized security state, at the moment of execution.
Continuity Seal extends runtime governance from authorization and policy enforcement into the integrity of the execution itself.
What a seal binds
| Binding | Verified at execution time |
|---|---|
| Workspace identity | The execution remains in the authorized workspace. |
| Agent identity | The same agent that was authorized is executing. |
| Execution identity | The execution matches the sealed execution identity. |
| Action, resource, tool | The action, target resource and tool have not changed. |
| Execution input | Exact input is verified after canonical normalization. |
| Authority state | Effective authority has not mutated or drifted since authorization. |
| Containment state | No containment condition is active. |
| Seal integrity | The seal itself passes cryptographic integrity validation. |
Authority integrity
- Authority-state fingerprinting
- Authority mutation detection
- Authority drift detection
- Transitive authority reconciliation
- Workspace-scoped authority verification
- Fail-closed execution when authority changes
Enforcement at the boundary
VALID STATE CHANGED STATE
| |
v v
ALLOW BLOCK
| |
v v
EXECUTION NO EXTERNAL SIDE EFFECTA continuity violation is enforced before the external side effect occurs. Modified action payloads, resources, tools or execution inputs are rejected when they no longer match the sealed execution state.
Provider and tool execution
Provider execution sealing, identity binding, runtime verification before invocation, input integrity and fail-closed enforcement.
ToolRegistry authorization binding, continuity-aware receipts, execution-time seal verification, exact input verification and authority mutation blocking.
Effect reconciliation
After execution, the runtime reconciles the observed effect against the sealed execution identity: authorized intent, sealed execution, actual effect. The reconciliation result is part of the evidence record.
Evidence and audit
Evidence is a first-class security object. Every decision leaves a record, including the ones that never executed.
- Canonical evidence creation at the enforcement boundary
- Persistent evidence storage
- Deterministic trace, decision, execution and evidence correlation
- Synchronized evidence finalization
- Retention for blocked and prevented actions
- Retrieval by trace, decision and agent
- Execution-level linkage through sealId and continuity enforcement status (v4.0)
A BLOCK is not silence. The prevented action is retained with its decision, so an auditor can see what was attempted and why it was stopped.
The enterprise layer adds an immutable evidence ledger and an enterprise audit ledger, with trace, decision, execution and evidence correlation preserved end to end.
Enterprise governance
The enterprise runtime introduced in v3.0 adds identity, access, registries and policy management around the core pipeline.
| Area | Capabilities |
|---|---|
| Identity and access | Workspace isolation, RBAC, server-side authorization, tenant boundary enforcement |
| Agent and provider governance | Agent registry and discovery, provider and model registry, governed model resolution, provider and model identity integrity |
| Policy and risk | Policy Studio, policy versioning, risk detection, deterministic decision precedence |
| Tool security | Tool security evaluation, ToolRegistry enforcement, external tool governance, MCP tool governance |
The runtime is designed around zero-trust execution, tenant isolation, auditable decisions and evidence-backed enforcement.
Approval engine
ESCALATE decisions stop execution until a person or process resolves them.
PENDING -> APPROVED
-> REJECTED
-> EXPIRED- Approvals are workspace-isolated.
- Approval state is enforced at runtime, not only recorded.
- Resolution is synchronized back to the governed execution.
Tool and MCP security
Tools are the point where an agent's intent becomes a side effect, so they are governed by the same pipeline as providers.
- Tool security evaluation before authorization
- ToolRegistry enforcement as the single canonical authority
- External tool governance
- MCP tool governance
- Authorization receipts bound to the Continuity Seal (v4.0)
- Post-execution effect reconciliation (v4.0)
Operations and entitlements
Operational surfaces that let a security or platform team run governance day to day.
Enterprise realtime events, a runtime-to-enterprise event bridge, persistent realtime writers and event subscriptions.
Enterprise incident ledger, governance orchestration, alert rules and webhook governance.
Usage ledger and metering, persistent usage projection, execution entitlements, provider-level entitlement and execution quota enforcement.
Billing state, SSO identity, SCIM provisioning and enterprise recovery contracts.
Packages
As of v2.0.0, nine packages are aligned to a single version.
| Package | Area |
|---|---|
@aegisora/core | Core types and shared foundations |
@aegisora/runtime | Governed execution runtime |
@aegisora/policy-engine | Policy evaluation |
@aegisora/security-engine | Security and risk analysis |
@aegisora/audit | Audit records |
@aegisora/storage | Persistence |
@aegisora/observability | Runtime observability |
@aegisora/plugins | Extension points |
@aegisora/sdk | Developer SDK |
Areas are summarized from package names and release notes. Consult each package README in the core repository for exact APIs.
Integrations
Aegisora intercepts agent actions rather than replacing the framework that produces them.
| Integration | Status |
|---|---|
| LangChain and LangGraph | Supported since v0.1.0 for runtime request interception |
| MCP tools | Governed through tool security evaluation (v3.0) |
| Model providers | Governed provider and model resolution; Anthropic is documented on docs.aegisora.org |
| External tools and APIs | Governed through ToolRegistry authorization |
Getting started
The smallest complete governance flow, and the rules for configuring it safely.
Agent -> Request -> Governance Runtime -> Decision -> ExecutionConfiguration
Configuration covers runtime behavior, policies, providers, logging and environment variables. Never commit API keys or other credentials; use local environment configuration for development secrets.
Maintained executable examples live in the examples directory of the core repository and run as a smoke test on every release.
Release validation
Release claims are tied to checks that run against the actual artifacts.
v4.0.0 validation, as published in the release notes
| Check | Result |
|---|---|
| Continuity Seal adversarial tests | 20/20 passed |
| Runtime continuity boundary tests | 5/5 passed |
| Runtime tests | 115/115 passed |
| Build, typecheck, runtime typecheck | Passed |
| Example smoke test | Passed |
| Security Readiness, CodeQL | Passed |
| Post-merge main validation | Passed |
| Provider boundary and ToolRegistry receipt validation | Passed |
| Authority and input mutation blocking, effect reconciliation | Validated |
Supply-chain discipline
- Security Readiness enforced in CI
- Package metadata, npm tarball content and package scope validation
- Release artifacts regenerated from the final release tree, with SHA256 evidence (v2.0.0)
- Release gate metadata rebound to the final commit
Release history
From a governed-execution foundation to execution continuity.
| Version | Date | Name | Focus | Commit |
|---|---|---|---|---|
| v4.0.0 | Sep 2026 | Continuity Seal | Execution continuity and runtime security | a99fa63 |
| v3.0.0 | Sep 10 | Enterprise Governance Runtime | Workspaces, RBAC, approvals, usage, entitlements | 338fbce |
| v2.0.0 | Aug 30 | Governance Platform | Nine aligned packages, unified platform | 08033e3 |
| v1.8.0 | Aug 30 | Runtime Hardening | Correlation identity integrity | 99da016 |
| v1.7.0 | Aug 30 | Governance Boundary | Boundary contract, fail-closed identities | 94521df |
| v1.6.0 | Aug 30 | Canonical Authority | ToolRegistry as single authority | 3090fd5 |
| v1.5.0 | Aug 29 | Evidence Integrity | Evidence as a first-class security object | e24aee3 |
| v0.1.4 | Aug 26 | Release Hardening | Security Readiness CI, package validation | 02e5479 |
| v0.1.2 | Aug 21 | Stabilization | Package organization, developer onboarding | 01a74b7 |
| v0.1.0 | Aug 18 | Foundation | First governed-execution path | a494e78 |
Each release is tagged in the core repository. Compare views and full notes are available on the GitHub releases page.