Microsoft makes its Copilot agent harness available through Foundry
Microsoft exposes the Copilot harness behind Copilot Code, Copilot Studio and the CLI through Foundry as a multi-model agent layer.
AI Daily Recap
34 articles · 4 categories
The finishable daily brief
What happened in AI — Oct 28, 2026
Wednesday, Oct 28, 2026
34 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Microsoft makes the multi-model Copilot harness available through Foundry.
- Cursor ships the Agent SDK for event-driven agents defined as code.
- 270 coding-agent runs test whether guardrails pay off.
- 324 agent sprints: 79% of committed stories done in-sprint, merge as slow as build.
- Google open-sources Agent Mesh for verifiable agent identity.
- Google DeepMind announces Gemini 4 Carbon.
The agent harness is turning into a shippable product: Microsoft opened the Copilot harness through Foundry and Cursor released its Agent SDK, while LangChain pushed scheduled and memory-backed agents.
The best evidence today is measurement: 270 guardrail runs and 324 agent sprints put numbers on what actually keeps coding agents reliable.
Prefer to watch?
LLM Digest Daily Explainer: Oct 28, 2026
Agent harnesses and runtimes are becoming products 4 items
Vendors are shipping the harness itself as a reusable layer: multi-model, event-driven, scheduled and memory-backed.
The Agent SDK
Cursor releases the Agent SDK it uses internally, for agents defined as code that act on events such as PR reviews or incidents.
How to Build Proactive, Always-On Agents with the Schedules SDK in Managed Deep Agents
LangChain's Schedules SDK for Managed Deep Agents frames proactive agents as cron-triggered runs.
How to build evidence-grounded agent memory with OpenWiki and Graphs
LangChain shows an evidence-grounded memory layer built on OpenWiki and graphs, so agents stop rebuilding context per question.
Measured results on coding-agent guardrails and process 4 items
Today's strongest posts report run counts and scores, not opinions, on what makes coding agents reliable.
Do Guardrails Make AI Coding Agents Better? What 270 Runs Showed
270 unattended runs across three models and seven tasks test whether instruction files and pre-commit gates pay for themselves.
Agent Scrum: Running a team of AI coding agents in two-hour sprints
Across 324 two-hour agent sprints, 79% of committed stories finished in-sprint, and approval-to-merge took as long as the build.
Getting an analytics agent from 50% to 96% accuracy
Benchouse took an analytics agent from 50% to 96% accuracy and found that more documentation sometimes made it drop the expected calculation.
Artificial Analysis v5 and the benchmarker's dilemma
Interconnects argues independent benchmarks are hard to keep meaningful at the frontier, framed around Artificial Analysis v5.
Agent identity and sandbox claims get testable 2 items
Two open-source projects target trust boundaries: verifiable agent identity and runnable contracts for sandbox claims.
Agentmesh – open-source zero-trust agent network from Google
Google's Agent Mesh gives each agent node a verifiable identity and a way to publish and call tools across a private network, including behind NAT.
I ran each AI coding-agent sandbox's own tests against the others
sandbox-conformance turns sandbox claims such as no key reads or no network into runnable contracts with per-sandbox results.
Frontier model and serving infrastructure 3 items
A new Google frontier model was announced, alongside serving-side moves on freshness and routing.
Gemini 4 Carbon: Our next frontier model
Google DeepMind announces Gemini 4 Carbon as its next frontier model; the announcement text in the feed carries no further detail.
Airbnb Cuts Recommendation Staleness from 2 Days to Seconds with Chronon
Airbnb used Chronon to cut recommendation feature staleness from two days to seconds.
Requesty: AI Gateway and LLM Router for 600 Models in EU Datacenters
Requesty offers one API over 600+ models with routing, failover, spend limits and per-request observability in EU datacenters.
You are caught up for this edition