Agent changelog digest: 5 features to adopt, 5 to try

Five changes from the last three months are worth adopting now: Claude Code subagent forking (CLAUDE_CODE_FORK_SUBAGENT) and the AGENTS.md fallback, Codex auto-reviewed approvals (approvals_reviewer) and managed worktrees (--worktree), and the OpenRouter Batch API (POST /api/v1/batches). Five more are worth a try. This is for developers who already run one coding agent and want to know what else is worth turning on.

The window is 2026-07-26 to 2026-10-26: Claude Code v2.1.221 to v2.1.310, Codex CLI 0.146.0 to 0.164.0, opencode v1.18.6 to v2.1.4, and OpenRouter’s dated announcements. Bug fixes, admin-only and plan-gated items are left out, and so are model launches such as the Claude 5.5 family and GPT-6.1 Sol (covered in Claude 5.5 migration). Scores are my judgement from the changelog text, not vendor data, and I did not run any of these features. The rubric: breadth of use, workflow impact, time or cost saved, maturity. 8 means adopt now, 6 to 7 means try it, 5 or lower is situational.

Five terms first

• Worktree: a second checkout of the same git repo in its own folder, so an agent can edit files without touching your working copy.
• Prompt cache: the provider keeps your already-processed context for a set time (5 minutes or 1 hour, depending on settings). Reading it back costs far less than sending it again. A pause longer than that lifetime means you pay full price to rebuild it.
• TTL: time to live, how long the provider keeps that cache.
• Compaction: the agent summarizing a long session to shrink its context.
• Subagent: a second agent session that the main agent starts to handle a side task. A fork is a subagent that starts as a copy of the main conversation instead of from a blank context.

The top 10

Claude Code

Subagent forking (8). A subagent with subagent_type: "fork" inherits the whole conversation and its prompt cache, so it starts warm. Since v2.1.232 subagent spawns in interactive sessions run in the background by default. Use /subtask to fork from inside a session. To turn it off, set "CLAUDE_CODE_FORK_SUBAGENT": "0" under env in settings.json. v2.1.307 (2026-10-21) tells parallel forks which launch is theirs so they stop starting each other’s work. Changelog.

The opt-out has its own note: turn off Claude Code subagent forking.

AGENTS.md fallback (8). In a project with no CLAUDE.md, Claude Code now reads AGENTS.md, so one instruction file can serve Claude Code and any other agent that reads AGENTS.md. Change the behavior under “Project instructions” in /config. If you keep both files they drift apart: edits land in one and the other goes stale, so pick one as the source of truth. Why agents ignore your CLAUDE.md covers the failure modes.

Prompt-cache visibility and TTL controls (7). /cost now shows a per-session cache line (hit ratio, misses, tokens re-cached) and a likely cause for a miss, such as changed tool definitions or an idle gap past the TTL. API-key and cloud-provider users can keep a 1-hour cache on the main conversation while subagents stay at 5 minutes.

Both values take 5m or 1h (note: prompt cache TTL split).

{
  "promptCacheTtl": "1h",
  "subagentPromptCacheTtl": "5m"
}

The PreModelSwitch and PostModelSwitch hook events (scripts you attach to a Claude Code event) can block or annotate a model switch. The related idle-compaction behavior is covered in Idle compaction trades your context for a cache write; the cache mechanics in Don’t break the prompt cache.

Honorable mentions at 6 to 7: /doctor prompt-audit for CLAUDE.md written for older models, /skill-doctor to find unused skills that cost context, claude plugin eval for scoring a plugin in CI, and claude --restricted plus --permission-prompts none for unattended runs.

Codex

Auto-reviewed approvals (8). Eligible approval prompts go to a reviewer subagent instead of you. The sandbox boundary is unchanged; the reviewer only replaces the human for requests that would cross it. Shipped in CLI 0.147.0. The documented config form:

codex --sandbox workspace-write --ask-for-approval on-request \
  -c approvals_reviewer=auto_review

The changelog also names a --approve-for-me flag, but I could not find it in the command docs, so treat the config form as the safe one. Docs.

Managed Git worktrees (8). Experimental in 0.154.0, on by default in 0.156.0 (2026-09-22). Isolated checkouts for new or forked sessions that you can browse and resume.

codex exec --worktree "implement the next step"
codex exec fork <SESSION_ID> --worktree "try an alternative approach"

--worktree needs local execution and cannot be combined with resume or review. Command reference.

Agent command center (7). codex agents opens a dashboard to search, start, rename and stop tasks. codex queue --thread <SESSION_ID> --message "Run the remaining checks" sends a message to an existing session. Agents can reference other tasks with @.

Breaking changes to act on

API-key sessions are not affected by the model retirements. New in the same window: /import brings instructions, skills, plugins and chats from Claude Code and Cursor into the Codex CLI.

opencode

OpenCode 2 is the active line. v1 and v2 both install the opencode command, so they cannot sit side by side; opencode upgrade --major (or /update) migrates. v1 stopped auto-updating, and v2.1.0 (2026-10-20) broke the SDK and plugin contracts.

Worktree tools (7). v2.0.27 added worktree_create, worktree_list and worktree_remove. Removal needs edit approval, and force: true can discard modified or untracked files.

Two features worth a look at 6, both aimed at the cache and context bill. Background compaction (v2.1.4, experimental) starts the summary at 85% of the compaction threshold so a session does not stall:

{ "experimental": { "background_compaction": true } }

Session warming sends periodic one-call requests to keep a provider cache alive while you pause. It is off by default and costs real tokens; I could not verify which release introduced it, only that the v2 docs list it:

{ "warming": true }

For the original opencode setup, see opencode as a free coding agent.

OpenRouter

Batch API (8). Thousands of requests in one call at typically about 50% of the synchronous per-token price. Limits: 200 MB per batch, one model and one endpoint per batch, results within 24 hours on most providers, data deleted after 30 days. It fits evals, classification, labeling and corpus embedding, not interactive work.

POST /api/v1/batches
endpoint: /v1/chat/completions | /v1/responses | /v1/messages | /v1/embeddings
model, requests[] (each with custom_id), optional callback_url

Poll GET /api/v1/batches/{id} until status reaches completed. Announcement.

Ori Harness and Ori Eval (7). Ori runs the agent CLI you already use on OpenRouter with any catalog model and one bill.

curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash
ori login
ori claude   # or: ori codex, ori opencode

Ori Eval turns “which model is best for my project” into an agent task that writes an eval, runs candidate models through one harness and reports scores and cost: ori code -p "What is the best model for my support agent?". Caveat: the Codex page cites Ori 0.15.6 while the public Ori changelog stops at 0.10.1, so check ori update output rather than trusting one page.

What the quarter shows

Across the four tools three themes recur: parallel agents isolated by worktrees or forks (Claude Code, Codex, opencode), less approval clicking (Codex), and prompt-cache cost controls (Claude Code, opencode). If you adopt one thing from each, take worktrees (Codex --worktree or opencode worktree tools), promptCacheTtl in Claude Code, and the OpenRouter Batch API for evals, labeling and other bulk jobs that can wait 24 hours.

Next step: run /cost in your next long Claude Code session and check the cache line before changing any TTL setting, then read Don’t break the prompt cache.