AstAstra AGI RuntimeEngineering case study
PRIVATE · AGENT RUNTIME

Astra AGI Runtime

A governed agent runtime where model-generated proposals do not receive authority over policy, persistence, sandboxing, execution or verification.

Private productv7.0.036 tests verifiedGoverned execution

The problem

Why this project exists

Agent runtimes need a control plane that remains authoritative when the model is uncertain, incorrect or compromised. Astra keeps planning separate from governed execution.

Verified capabilities

What is actually implemented

These capability descriptions are intentionally limited to the evidence-backed implementation surface.

01

Durable state

SQLite persists runs, nodes and checkpoints.

02

Persistent memory

Memory carries provenance, confidence, utility, relevance and decay.

03

Governance

Risk, authorization and approval gates remain outside worker logic.

04

Sandbox

Shell and Python execution cross an explicit sandbox boundary.

05

Evidence receipts

Canonical serialization and SHA-256 chaining create tamper-evident receipts.

06

Self-improvement gates

Candidate changes require tests, benchmarks, security, regression and approval before promotion.

Architecture

How the system is structured

The diagram below represents the current verified architecture boundary; planned or environment-dependent components are described in the engineering notes rather than silently promoted.

Perceive
↓
Understand
↓
Recall
↓
Reason
↓
Plan
↓
Govern
↓
Act
↓
Observe
↓
Verify
BOUNDARY

Evidence and authority

Approval-required actions remain blocked without an explicit decision.

Cross-session memory retrieval and receipt reload verification are regression-tested.

Engineering deep dive

Decisions, trade-offs and boundaries

Technology names matter less than the engineering contracts they enforce.

E1

Approval-required actions remain blocked without an explicit decision.

E2

Verification failures are represented as failures rather than unconditional success.

E3

Sandbox execution validates arguments, scopes paths, limits output and enforces timeouts.

E4

Cross-session memory retrieval and receipt reload verification are regression-tested.

Project status

Private · evidence-backed status

Repair verified: 36 full-suite tests passed and 12 focused repair regressions passed. Remaining distributed and irreversible-operation limitations are documented.

ClassificationPrivate product / engineering case study
EvidenceRepository implementation, tests and project records cross-checked
ClaimsCurrent, planned and limitations kept separate
DeploymentNot changed by portfolio integration
PRIVATE PRODUCT

Architecture can be discussed without exposing the implementation.

Private implementation — verified repair evidence and architecture can be discussed.