Architecture AI Updates: October 27, 2026

1. Chains, Graphs, Code: LinkedIn’s Three Generations of Agent Harness

InfoQ. In a new article for InfoQ’s Agentic AI Architecture series, LinkedIn’s Karthik Ramgopal divides agent harnesses into three paradigms. Chains are linear prompt sequences. They are cheap and fast, but one upstream error cascades through the flow, as happened in LinkedIn’s early three-step route, retrieve and generate assistant. Graphs make control flow explicit with nodes and edges, adding conditional routing, retries, state and validation. LinkedIn built Hiring Assistant on LangGraph this way and reused its messaging system for FIFO agent-to-agent delivery, but graphs became brittle as recovery paths piled up. Code harnesses treat execution as the primitive: the agent writes SQL and Python, runs it in a sandbox against sample data, checks it against privacy and governance constraints, and loops until it passes. LinkedIn uses this for an internal analytics agent. The author’s rules apply at every stage: treat LLM output as untrusted input, keep planning separate from execution, invest in observability, and pick the simplest abstraction that meets the reliability, speed and cost targets. Source

2. Salesforce Takes Rules, Identity and Context Out of the Agent Prompt

ByteByteGo. ByteByteGo describes how Salesforce redesigned Agentforce after its first version, which put every instruction in one large prompt, started skipping rules once there were more than about eight. The new design moves four jobs out of the prompt and gives each to a dedicated system. Agent Script enforces rules deterministically on every turn. Identity resolution in the Data 360 layer merges records from many systems into one customer profile. An Agent Context Engine decides which data reaches the model. Agent Identity gives each agent its own permissions. The model only handles the steps that actually need reasoning. The article cites the IFScale benchmark, where even the best of 20 models followed only 68% of 500 instructions, as evidence that a prompt can shape behavior but cannot enforce it. Source

3. Per-Request Model Routing, and Why Risk Labels Should Persist Across Turns

InfoQ. InfoQ looks at the routing design behind Cloudflare’s Auto Router, now in public beta in AI Gateway. It works in two stages. First it filters a candidate pool by format support, credentials, access control and spend limits. Then a classifier rates each request from 1 to 5 on complexity, ambiguity, stakes and dependence on earlier context, and a scoring matrix trades expected quality against an adaptive cost penalty. The classifier reads only the newest turns, capped at 8 KiB. Practitioners quoted in the piece point out that this lets a short “go ahead” that commits a risky action look low-stakes, and they recommend carrying an application-owned risk label across the whole session. Other design points: routing decisions are exposed in response headers so they can be audited, the pool can be restricted per request, and session IDs matter because switching models mid-conversation loses the prompt cache. Source