Astra · Agent Runtime

The Agent Runtime for Enterprise Work

Agents that don't just answer once, but keep working toward a goal: understand the objective, build a trackable task graph, call governed capabilities, keep the evidence, resume after waiting, failure or disconnection — and explain the whole thing afterwards.

1
Shared runtime backbone
4
Capability providers
0
Work lost on disconnect
100%
Traceable & auditable
ASTRA
Runtime
Context
Prompt
LLM
Verify
Memory
ONE BACKBONE · MANY PROVIDERS

One agent runtime backbone, many capability providers

Web, CLI, SDK, enterprise apps and edge executors don't each implement their own agent. They share the same session, run, Work, context and audit semantics — they differ only in what capabilities they can reach.

Interaction surfaces
WebCLI / TUISDKEnterprise appsEdge

Shared runtime backbone

SessionRunTurnWorkTask graph
ContextPolicyCheckpointRecoveryAudit
Capability providers

Cloud capabilities

Sessions, memory, tasks, reports and governed external access

Cloud services

Edge executor

Local files, shell, Git, browser and private networks

User device / workstation

MCP capabilities

Knowledge bases, databases, ticketing and approval systems

MCP service environment

Sandbox executor

Workloads needing isolation, quotas or a specific runtime

Sandbox runtime

Clients are different entrances, providers are different hands — the thinking, the state and the recovery always live in one backbone.

What that backbone buys you

Continuity

Continuity

Work doesn't depend on a page or a process. It pauses, resumes and follows you across devices.

Control

Control

Capability, permission, network, data and side-effect boundaries are explicit — block one action precisely without killing the whole task.

Explainability

Explainability

Not just the answer: which context was used, which tools were called, what failed along the way, and why this route was chosen.

RESEARCH

Context assembly as a systems problem

The same problem can be met by tuning prompts from experience, or treated as a systems problem that can be optimized and measured. Astra takes the second path.

VLDB 2026arXiv:2609.00749

ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents

Peng Xu, Zuyu Zhang, Yuze Sun, Feng Tian, Long Wang, Chen Zhang

Every model call in a long-horizon agent has to answer the same question: what goes into this prompt, in what order, and when does history get compacted. And it has to answer under two hard constraints — a fixed context-window budget, and a byte-sensitive prompt cache where changing one byte can invalidate the whole segment. The paper brings database query execution to that problem: data sources are catalogued like tables, assembly runs as a five-phase pipeline, and the optimization is deterministic rather than left to the model.

31%
fewer tokens
23%
fewer LLM calls
9%
faster response

Measured on the Qutebrowser subset of SWE-bench Pro, against append-only context assembly.

Read the paper
FROM ANSWERS TO LONG-RUNNING WORK

From a single answer to work that keeps going

This is the line between an agent platform and a chat endpoint.

Work
Task graph
Checkpoint
Evidence
Approval
Recovery
Delivery

One question usually takes one model call. Real work doesn't: analysing code, debugging an incident, working a ticket queue, running a deployment, producing a report — all of it takes many steps and many tools. Once work gets long, the hard parts show up. One HTTP request can't cover the whole execution. Should the task continue after the UI disconnects? Retry, reroute or stop when a tool fails? How does a human approval get back into the original task? How does this turn's evidence reach the next one? Astra makes that state explicit and durable — part of the runtime, instead of scattered across UI, scripts and prompts.

“Long-running work doesn't need a smarter prompt. It needs an explicit, durable, resumable runtime.”
01

Progress and activity

Work and the task graph make a long goal visible: what's done, what's running, what's waiting. You always know what the agent is doing.

02

Durable intermediate results

Every step's output and evidence is kept, so the next turn can rely on it instead of asking the model to remember.

03

Resumable checkpoints

Runs checkpoint at meaningful boundaries. After a failure, a wait or a disconnect, work continues from there — not from scratch.

04

Cancel, pause, continue

A run is a controllable execution: cancel it, pause it, or let it continue once a condition is met. The boundaries are explicit.

05

Input and approval routing

When a human is needed the run enters an explicit wait, and the answer returns to the original task rather than starting a new thread.

06

Background execution and delivery

Work keeps moving after the connection drops, then delivers a result or a failure reason — with traceable evidence and a next step.

CLOUD BRAIN · LOCAL HANDS

Cloud brain, local hands

The cloud is the right place for shared state, model calls and orchestration. Your files, code, browser sessions and private network usually are not. The edge executor lets an agent work in your workspace without giving up your local data boundary.

Cloud runtime backbone

Shared facts and coordination

Session / run / Work
Context and memory
Model routing and quota
Policy, approval, audit
Cross-device sync

Edge executor

Hands inside an authorised boundary

Bound local workspace
Files and shell
Git / build / test
Local browser
Private network and local MCP

The edge executor does not run a second agent loop. The cloud holds the control facts for sessions, runs and Work; the edge only executes local work inside a bound, authorised boundary.

What happens when the link drops

Local files and shell go 'provider offline' — the whole agent doesn't die with them
Cloud-side memory, sessions, tasks, plans and reports stay available
The current run can wait for reconnect, or take an explicit fallback per policy
Results produced but not yet uploaded queue for sync
You see sync progress, failure reasons and what to do next
SIDE-EFFECT TIERS

Permissions decided by what a call actually changes

A side effect isn't judged by the tool's name, but by what this call really changes, who it affects, whether it can be rolled back, and whether it crosses a trust boundary. The runtime then decides: run it, ask first, allow it only as a suggestion in planning mode, or block it.

TierExamplesTypical policy
Read-onlyRead a file, check status, query a reportAllowed if provider and user permissions permit
Local changeWrite files, run tests, modify GitRequires a bound workspace and approval
External readBrowse the web, call a business APISubject to network and data policy
External changeFile a ticket, publish, write to a databaseExplicit approval or a higher trust level
High riskDeletion, credential exposure, dangerous network opsDenied by default or hard-gated

Planning mode: think first, then act

Planning mode lets the agent understand the problem, read context, break down steps and propose a plan — while side effects like writing files, changing code or sending external messages stay blocked by default. It's a change of execution policy, not a second agent.

Why Runtime, Not Framework

Frameworks give you parts. Runtimes give you guarantees.

DimensionFramework (LangChain / CrewAI)Coding Assistant (Cursor / Claude Code)Astra Agentic Runtime
FormA pile of libraries, DIYSingle-purpose productManaged runtime + guarantees
What you writeWire everything from scratchUse someone else's toolJust a system prompt + skills
Decision auditNoneNoneEvery decision bound to a snapshot
Memory managementRAG bolted onNone5-layer cognitive architecture, self-curating
Production experimentsGamble in prodN/AClone prod DB, zero-risk regression
Cost & trustEach piece separateBlack boxUnified cost + confidence signals
“JVM is to Java what Astra is to Agents.”

Six Platform Capabilities

Agent developers write prompts and pick skills — the runtime does the rest.

Decision Audit

Decision Lineage — every decision replayable
  • Decision = f(prompt@version, skill@version, context@snapshot, memory@state)
  • Given a decision_id, reconstruct exactly what the LLM saw — 100%
  • Causal-chain event stream across every session for compliance & audit

Cognitive Memory

Cognitive Memory — not a RAG afterthought
  • Five-layer model: sensory / working / episodic / semantic / procedural
  • Tabular + graph backends; vector + fulltext + SQL in one query
  • Self-curating: confidence decay, quarantine, compaction

Self-Evolving Loop

Self-Evolving — no more 'ship and pray'
  • Implicit feedback mining → LLM diagnosis → auto-improve → regression gate → activate
  • Clone-Test-Merge on real data; must pass the gate to ship
  • Auto-rollback if regression fails — iteration becomes engineering

Trust & Safety

Trust & Safety — trust is a platform primitive
  • Hallucination Firewall verifies every LLM claim against the snapshot
  • Tool Result Quality Firewall assesses tool outputs before they enter the LLM
  • Edge/Cloud provenance tags; every answer carries a confidence signal

Skills as Packages

Skills — with schema, migrations and distribution
  • Platform-defined schemas using `sk_{skill}_{table}` naming convention
  • MCP-compatible protocol with a typed API layer on top
  • Progressive disclosure — skills load by task type

Zero-Risk Sandbox

Git-for-Data — safe trials on production data
  • CREATE CLONE: millisecond-level full-DB copies for experiments
  • Time Travel: reconstruct state at any past point in time
  • Branch / Diff / Merge: Git-like workflow for your data

Layered Architecture

Edge runs tools, cloud owns LLM and state, MatrixOne powers enhanced services.

EDGE · User Machine
EdgeChatLoop · local tools · MCP servers · permissions — tools run on the user side, state syncs to cloud
USER AGENTS · Business Agents
Code Review · CI Diagnosis · Data Analysis · Custom agents — one AgentProfile = system_prompt + skill set + model
SYSTEM AGENTS · Platform Guardians
Regression · Audit · Tuning · Eval — same execution model, elevated privileges, auto-triggered
PLATFORM SERVICES
Memory · Context · Skills · Planning · Trust Engine · LLM Router · Streaming · Evaluation · Cost Control
PLATFORM STATE · MatrixOne
Identity · sessions · events · snapshots · audit · skill catalog · credentials — all in one MatrixOne
ENHANCED SERVICES
Sandbox clone · time travel · hybrid search · branch/diff/merge · skill marketplace · dynamic table — auto-unlocked on MatrixOne

Adding a new business agent = define a system prompt + pick skills + pick a model. Zero platform code changes.

What Happens in One Agent Turn

Three steps that reveal Astra's runtime guarantees.

01

Context Assembly

Walk the causal chain → hybrid-search the memory → load skill definitions → allocate token budget by task type. Every token has to earn its place.

02

Snapshot + LLM Call

Before the LLM call, asynchronously write the context snapshot. Model, parameters and input are all frozen for later audit and replay.

03

Verify + Evolve

Hallucination Firewall verifies each claim → decision lands in audit log → implicit feedback flows back → regression gate drives the next round of prompt and skill evolution.

Replay Any Turn, At Any Time
user.msg
context.snapshot
skill.load
llm.call
firewall.verify
decision.log
feedback.in

Every step written into the event stream and snapshots — ask 'why did the agent do that' three months later, and get the exact answer in seconds.

Developer view — just a few commands.

Use Cases

From customer-facing assistants to platform-internal guardian agents.

Business

Intelligent Support & Enterprise Knowledge

Every reply traces back to the product doc version cited at the time. When docs update, impacted past answers auto-enter QA review — no more serving stale information to customers.

Business

Sales Assistant & Contract Review

The agent generates recommendations from customer history, contract libraries and compliance rules. Each suggestion carries a confidence score so reviewers can focus only on low-confidence sections — faster without losing control.

Business

Data Analysis & BI Copilot

Business users ask questions in natural language. The agent runs analysis on a zero-copy clone of production — heavy queries don't impact live workloads. Results bind to data snapshots so you can recreate the exact state later.

Business

Operations & Marketing Assistant

When recommending segmentation or campaign strategies, the agent runs four approaches in parallel in sandboxed clones, computes expected ROI from historical data, then suggests the best. Decisions are grounded and reviewable.

Technical

Code Review Agent

Every suggestion binds to the code snapshot and repo state at the time. Three months later, you can reconstruct the full code context of any past suggestion — no more 'the AI said something but we can't reproduce it.'

Technical

CI / Ops Diagnostic Agent

Runs as a System Agent triggered by failures: pull context → reproduce in a cloned environment → propose a fix. Every diagnostic step is recorded — on-call has receipts.

Stack & Deployment

Rust-first performance core, MatrixOne as platform state store.

Stack

  • Rust + Axum + Tokio async event pipeline
  • MatrixOne as platform state (vector + fulltext + HTAP + time travel)
  • Memoria as the cognitive memory backend (also runs standalone)
  • Multi-provider LLM routing with circuit breakers and fallback chains

Deployment

  • Local dev: `make dev-start` — full stack in 5 minutes
  • Docker All-in-One: dependencies + app in one command
  • Kubernetes: edge/cloud split, production-grade
  • GPU optional with Ray scheduling for high concurrency

Proof from customers

Digital TV / Smart Devices

Smart Home IoT Leader Evaluates RFQ Requirements and Fills the Original Template

A Leading Smart Home IoT Company

A capability knowledge base and RFQ parser let an Agent classify each requirement as compliant, non-compliant, or uncertain, cite evidence, and fill the original Excel template.

3 states
Compliance decisions
Multi-sheet
RFQ parsing
Consumer Goods / Legal & Compliance

Amway: Contract Review From 16 Days to One Minute

Amway

A legal team at a multi-billion-dollar multinational handles a large volume of cross-border contracts, with long manual review cycles, risk findings that could not be traced, and standards that varied by reviewer. Amway built a unified legal knowledge brain on MOI that flags risk clauses, missing clauses and inconsistent definitions in seconds.

1 minute
Review cycle (was 16 days)
99.9%
AI risk identification coverage
Private Equity / Investment Research

An Industry Fund: From Artisanal Research to Industrial Decisions

An industry fund

Industry scanning, target identification and post-investment tracking all ran on manual effort, with internal and external data siloed apart. The fund built an AI investment intelligence system on MOI: multi-source data is collected and structured automatically, linked into industry-chain, company and technology knowledge graphs, with reports and alerts generated on top.

80%
Analyst time freed
100%
Static reports to live updates
Education / K-12 Learning

Tencent AI Tutor: Six Agents Collaborating on Structured K-12 Reading

Tencent

Students don't know what to read or how; parents can't see progress; teachers can't follow every child. Tencent built AI Tutor on MOI, using a multi-agent architecture that turns one reading session into a personalised path, interactive Q&A, comprehension testing and knowledge expansion.

6 agents
Collaborating per session
Personalised
Path per student level
Telecom / Procurement

China Mobile IoT AI Sourcing: From Procurement Files to Traceable Reports

China Mobile IoT

MOI connects procurement document parsing, item confirmation, multi-source research, supplier recommendations, price analysis, and report export in one evidence-backed workflow.

Multi-format
Procurement inputs
Internal+External
Source coverage
New Energy / Manufacturing

AI Bid Intelligence Agent: MatrixOne Intelligence Powers Enterprise Smart Bidding

Jinpan Technology

A leading new energy manufacturer built an AI data platform with MatrixOne Intelligence, enabling end-to-end automated bid document generation and improving bidding decision efficiency by 40%.

95%
Key-info extraction accuracy (was < 50%)
5 min
Cross-department evidence lookup (was days)

Bring Your Agents to Production

Astra already powers multiple business agents running in production inside MatrixOrigin. Contact us for a detailed solution or book a deep-dive demo.