Latest — AI & agent coding
Advanced Rolling digest of what's genuinely new in agent-assisted coding. Short summaries here, depth in the linked pages.Newest changes across Claude Code, Codex, Antigravity and Grok: Agent CLI game changers, checked November 2026.
This digest was last swept in August 2026. Features move fast — every claim below links to official docs; re-verify on your installed version before relying on it.
TL;DR — what changed recently
| New thing | One-liner | Where |
|---|---|---|
/goal | Keep working until a condition is verifiably true | Claude Code |
Cloud routines (/schedule) | Scheduled / GitHub-triggered cloud agents, no machine needed | Claude Code |
/loop dynamic | Recurring prompts; Claude picks the polling interval | Claude Code |
| Agent teams | Peers that message each other + shared task list (experimental) | Claude Code |
| Workflow size caps | workflowSizeGuideline, 16 concurrent / 1,000 agents per run | Claude Code |
Fast mode (/fast) | Opus up to ~2.5× faster at premium pricing | Claude Code |
| Auto memory | Claude persists per-project notes across sessions | Claude Code |
/usage | Breakdown of what's burning your quota | Claude Code |
Autonomy: run until done
The biggest 2026 shift: sessions that keep going without you watching.
/goal — condition-driven runs
/goal all tests in test/auth pass and lint is cleanA small evaluator model (Haiku by default) checks the condition after each turn; Claude continues until it holds. The goal survives --resume. Pair with --permission-mode auto for unattended runs — see auto mode.
Official: goal docs
/loop — recurring prompts, smart intervals
/loop check if the deploy finished # Claude picks the interval (1min–1hr)
/loop 5m run the flaky test and log result # fixed cron-style intervalEsc pauses a fixed loop; dynamic loops can end themselves when the watched thing completes. For long watches Claude may switch to a background Monitor instead of polling.
Official: scheduled tasks
Cloud routines — agents on a schedule, no laptop
Anthropic-hosted sessions triggered by cron schedules, GitHub events (PR opened/merged, releases), or an API webhook:
/schedule daily PR review at 9am
/schedule list
/schedule runConfigure network access, env vars, and MCP connectors per routine; manage in the CLI or at claude.ai/code/routines. One-off runs ("tomorrow at 9am, summarize yesterday's PRs") are rolling out.
Official: routines docs
Multi-agent: what moved
Baseline concepts live on Claude agents and Claude workflows; this is the delta.
Agent teams (experimental)
Teams got a real switch and real guardrails:
- Enable with
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1(off by default) - Teammates share a task list and message each other directly — the core difference from subagents, which only report back to the lead
- Display: in-process (default) or split panes in tmux / iTerm2
- Safety (July 2026): teammate permission prompts bubble up to the lead; a teammate can't relay an approval to bypass checks
Official: agent teams
Dynamic workflows: caps, sizing, reuse
- Hard caps: 16 concurrent agents, 1,000 per run
workflowSizeGuidelinesetting (small/medium/large/unrestricted) — defaultmediumkeeps runs under ~15 agents; shown in/configon recent builds- Save any run's script as a reusable
/workflow-namecommand - Bundled
/deep-researchworkflow for multi-source investigation with cross-checking
Official: workflows · context: Claude workflows
Subagent persistent memory
Subagents can keep their own memory across invocations — add to the definition frontmatter:
---
name: code-reviewer
memory: true
---Official: sub-agents
Speed, cost, visibility
Fast mode — /fast
Opus (5 / 4.8 only) at up to ~2.5× output speed, premium pricing (~$10/$50 per MTok), separate rate-limit pool with auto-fallback to standard. For live debugging and rapid iteration, not bulk runs.
Official: fast mode
/usage — see what's burning quota
Breaks usage down by parallel sessions, subagents, cache misses, and long context; toggle day/week views. Pairs with /cost and --max-budget-usd from the advanced page.
/effort slider
/effort with no arguments now opens an arrow-key picker across low → medium → high → xhigh → ultracode. Level details: ultra family.
Memory
Claude Code now auto-writes project memory — persistent notes distilled from your corrections and preferences, stored per project (MEMORY.md + topic files) and loaded each session. Toggle via /memory or the autoMemoryEnabled setting.
Official: memory docs
Small but handy
| Command | What |
|---|---|
/recap | One-line summary of what happened while you were away |
/fewer-permission-prompts | Scans transcripts, proposes an allowlist for .claude/settings.json |
PreCompact hook | Block context compaction ({"decision":"block"}) |
| Push notifications | Phone ping when Claude needs a decision (needs Remote Control) |
Other agent CLIs
This digest is Claude-Code-heavy. For the rest of the stack, start with Agent CLI game changers, then the section pages:
Related on this site
- Claude Code 2026 map — how agents / workflows / ultra fit together
- Claude Code advanced — permission modes, auto mode, sandbox, CI
- Claude agents · Claude workflows · Claude ultra