AI Architecture Updates: September 26, 2026

1. An Agent Is a Model Plus a Harness, and There Are Two Ways to Build the Harness

Trista Pan, writing for InfoQ, defines the agent harness as everything built around the model to turn it into a product, split into a development half (memory, tools, retrieval, prompts, orchestration) and an operations half (observability, evaluation, guardrails, routing, cost monitoring, scaling). She contrasts harness-as-a-service, where platforms like AWS AgentCore, Vertex AI, or Azure AI Foundry expose these components as managed APIs, with a self-managed harness built from SDKs and open-source parts on Kubernetes behind a gateway such as Agent Router (formerly Envoy AI Gateway). Her point is that both deliver the same capabilities, and the choice is really about who operates the infrastructure and how much portability the team needs. Source

2. Stateless MCP Lets Remote Servers Sit Behind Ordinary Load Balancers

InfoQ covered the MCP specification changes that make the protocol stateless, removing the initialize handshake and the Mcp-Session-Id header so any request can be routed to any server instance without sticky sessions or shared session storage. An optional server/discover call replaces up-front capability negotiation, a multi-request pattern replaces server-initiated requests over held-open streams, and new Mcp-Method and Mcp-Name headers plus W3C Trace Context support gateway routing, throttling, and distributed tracing. The practical result is that serverless targets like AWS Lambda become viable for MCP servers, though AWS recommends keeping legacy session infrastructure until older clients are gone. Source

3. Perplexity Replaced DynamoDB With a Purpose-Built Serving Layer for Its Answer Engine

InfoQ reported on CobbleDB, a Rust key-value store of roughly 40,000 lines that Perplexity built to serve web metadata, passages, and vector embeddings for its answer engine. It sits in a three-tier design: YTsaurus holds durable versioned state, a stateless consumer called Lorry batches exports into partition-aligned files on S3, and CobbleDB serves batched lookups from RocksDB on local NVMe with three replicas per partition and a router using zone affinity and speculative request hedging. Median batch-read latency fell from 31.4ms to 5.6ms and P99 from 123ms to 24.2ms, with storage costs down at least 20%. Source