Multi-Agent SystemsAug 2026 – PresentSoftware Engineer · Multi-Agent Architecture

QANTUM

Slack-Native Multi-Agent QA Orchestrator

A Slack-native multi-agent QA orchestration platform routing engineering and QA requests across specialized agents via an extensible capability protocol. Features shared and user memory, Model Context Protocol (MCP) integrations, and autonomous MR/PR generation and validation flows—backed by production usage across ~100 engineers.

100Engineers
5,199Agent runs
1,430Slack threads
0.17%Run failure rate
645MTokens
~$1,010LLM cost (~$0.19/run)
10.7sMedian latency
+62% / +47%MoM runs / users
The Problem & Engineering Constraint

The Core Challenge

Engineering teams interact with fragmented internal tools for test runs, device logs, bug triage, and PR validations. QA coordination between humans and automated tooling lacked context and unified execution.
Technical Architecture & Approach

Engineering Solution & Implementation

Architected a Slack-native orchestrator that routes user intents to specialized subagents via capability protocols and MCP tool registries, with dynamic model routing and persistent semantic memory. Production usage since Aug 2026 backs the architecture: thousands of agent runs, Slack-thread adoption, and gated confirmation flows where awaiting approval is not counted as failure.

SYSTEM ARCHITECTURE: QANTUM
MULTI-AGENT CAPABILITY PROTOCOL
SLACK / API CLIENT• /qantum slash cmds• Natural language QA• Fleet status triggers• Test run invocationsEntra SSO VerifiedCENTRAL ORCHESTRATORModel & Intent RouterOpus planning / light tacticalCapability Registry (Protocol)Dynamic tool discovery & MCPShared & User MemoryIsolated state & user preferencesEXPLORER WORKER• MCP Tool Ingestion• Device screen queriesFORGE WORKER• Test authoring triggers• Git worktree verificationVALIDATOR / REVIEWER• Isolated check memory• Automated MR commentsINFRASTRUCTUREDevice FleetsPhysical testbedsGit RepositoriesBranch & worktreesCI PipelinesAutomated gatingJira / Confluence

Figure 2: QANTUM's multi-agent capability protocol routing tasks between Slack and specialized agent workers.

QANTUM demo — Slack-native multi-agent QA orchestration
Measured Production Impact

Verified Outcomes & Deliverables

Adopted by ~100 engineers with 5,199 agent runs and 1,430 Slack threads in roughly eight weeks (floor from Aug 2026).

671 capability runs and 44 agent-authored merge requests in the same window.

0.17% run failure rate; confirmation-gate awaiting-approval states are not treated as failures.

Median run latency 10.7s at ~$0.19/run (~$1,010 LLM spend, 645M tokens), with MoM growth of +62% runs and +47% users.

Technologies & Components

System Tooling & Technologies

PythonClaude Agent SDKMCP ProtocolLiteLLMSlack BoltAlembicPostgreSQLDocker