What Is Agent Memory? How AI Agents Remember Your Work
Agent memory lets AI agents carry notes, preferences and decisions into later sessions. How it works in Claude Code, Codex and GitHub Copilot.

Agent memory lets AI agents carry notes, preferences and decisions into later sessions. How it works in Claude Code, Codex and GitHub Copilot.

An agent harness is the software around an AI model that runs it in a loop, gives it tools and manages what it sees. Here is what each part does.

Subagents are helpers an AI agent hands a side task to. Here is how they work in Claude Code, Cursor, Codex and Copilot, and what they cost in tokens.

A cloud coding agent writes and tests code on a remote computer, then hands back changes for review. How Cursor, Codex, Copilot and Claude Code do it.

We ran 8 AI-written jq commands on messy JSON. All 4 from a plain prompt crashed on one quoted number; a 5-rule prompt worked 4 out of 4 times.

VS Code 1.146 deprecates the Local agent harness, with removal set for February 2027, and adds an agent inbox, chat windows and Skills over MCP.

We ran 32 AI-written cron schedules on a fake clock. The model got 31 right, but one plain-prompt miss ran a monthly job 34 times instead of 4.

We ran 8 AI-written bash scripts on a tricky test folder. None lost a file, but plain prompts moved files at once and 3 of 4 hid a sorting bug.

We asked AI for a ZIP code regex 8 times. Every pattern worked in Python, but 2 of 4 plain answers broke in Google Sheets. Naming the engine fixed it.

Google Antigravity's agent now drives the app's built-in browser on personal accounts, with no Chrome setup, parallel tabs and approval gates for uploads.

Atuin Threads syncs Claude Code and Codex sessions across machines. HN fans liked it, security-minded developers balked, and many had built their own.

Google's /swarm runs teams of coding agents that burn quota faster than chat, and Anthropic data puts multi-agent use at about 15x. Use swarms for wide jobs.

We asked AI for the same SQL query 8 times. Without the schema, 0 of 4 ran. With schema and sample values, 4 of 4 ran and returned the right answer.

Anthropic's Lydia Hallie says Claude Code users should trim CLAUDE.md, let auto mode approve, and use deny rules, sandboxes and hooks for hard limits.

Google Antigravity is previewing /swarm, a command that splits one prompt across named agents on a shared board. Docs say it is on Google AI Pro and Ultra.

Use AI in Blender by script or by connector. We ran 8 AI-written scripts in Blender 5.3: all worked, but plain requests added things we never asked for.

Go from a fact sheet to a live landing page in 6 steps, with a copy prompt we tested on 8 calls that kept AI from adding promises you never made.

A Nov. 10 a16z essay says agents now make up the majority of web traffic. The essay cites no data for that claim.

Cursor Projects is now multiplayer in beta: teammates share a task board and context, and their coordinator agents can message each other directly.

AWS says AI agents wrote every line of AWS Emulator. Hacker News and X react: trust in unreviewed code, fidelity gaps, no security tests, telemetry, LocalStack.
