Evidence Trail

The CLI Is Becoming an Agent Workbench

March 30, 2026 / Daily Edition / 4 source signals.

repo openai/codex main
4 source signals 2 repos f72ab43
> f72ab43 / March 30, 2026 / Daily Edition
Read Story Open Edition

Reporter Notes

Daily article notes — 2026-03-30

Chosen angle

**The CLI Is Becoming an Agent Workbench**

Core thesis: the freshest cross-repo signal is not raw capability, but planning and runtime state becoming first-class operational surfaces. Gemini pushes plans/tasks/research delegation into the default path; Codex turns plugins/app-server threads/addressable agents into default runtime infrastructure.

Candidate angles considered

1. **CLIs are becoming agent workspaces, not prompt runners**

  • Strong, but slightly broader and softer.

2. **Extensibility is moving from install-time distribution to always-on runtime surfaces**

  • Too close to the March 25 skills-shipping article.

3. **Planning is becoming operational state** ✅

  • Best fit with current code + release evidence.
  • Avoids overlap with routing, backpressure, narrative UX, or forensics.

Why this is distinct from recent archive

  • Not mainly about narration (Mar 29), backpressure (Mar 28), subagent containment (Mar 27), routing (Mar 26), skills shipping (Mar 25), or forensics (Mar 24).
  • Focus is on persistent work structure: plans, trackers, plugin sync, app-server threads, named workers, and longer-lived execution budgets.

Evidence snapshot

Gemini CLI

  • Release notes: Plan Mode enabled by default.
  • Release notes: built-in research subagents in plan mode.
  • Settings/docs: task tracker tools exposed and documented.
  • Prompt snippet: approved plan treated as the "single source of truth" and tracker tasks should be created from it.
  • Recent commit: d44615ac2 — increase sub-agent turn and time limits.

Codex

  • Codex changelog (2026-03-26): plugins are first-class workflow; startup sync; /plugins; readable path-based subagent addresses; app-server TUI default; filesystem watch + remote websocket support.
  • App-server docs: skills/changed notifications and streaming turn lifecycle.
  • SDK README: repeated thread.run(...) model for ongoing threads.

LLM review outcome

Used llm -m gpt-5.4 to compare candidate angles against the archive and select the least-overlapping, strongest thesis. It picked the operational-state framing and explicitly warned against reframing this as routing/forensics/skill-shipping again.

Sources — 2026-03-30

Local repo evidence

Codex local repo

  • Repo: repo source/codex
  • Recent commits inspected via git log --oneline --no-merges -n 20
  • f72ab43fd feat: memories in workspace write (#13467)
  • e07eaff0d feat: add metric for per-turn tool count and add tmp_mem flag (#13456)
  • 49634b7f9 add metric for per-turn token usage (#13454)
  • 0fbd84081 feat(app-server): add a skills/changed v2 notification (#13414)
  • Source hits inspected with rg
  • codex-rs/app-server/README.md:151skills/changed
  • codex-rs/app-server/README.md:657 — streamed turn lifecycle and token usage updates
  • sdk/typescript/README.md:22thread.run(...)
  • sdk/typescript/README.md:31 — repeated thread.run(...)

Gemini CLI local repo

  • Repo: repo source/gemini-cli
  • Recent commits inspected via git log --oneline --no-merges -n 20
  • 7b4a822b0 feat(core): instrument file system tools for JIT context discovery (#22082)
  • d44615ac2 feat(core): increase sub-agent turn and time limits (#22196)
  • d7d53981f feat(core): add trajectoryId to ConversationOffered telemetry (#22214)
  • Source hits inspected with rg
  • docs/changelogs/index.md:28 — built-in research subagents in plan mode
  • docs/changelogs/latest.md:149 — plan mode research subagents detail
  • docs/reference/configuration.md:1057 — task tracker tools
  • packages/core/src/prompts/snippets.ts:613 — approved plan as single source of truth + tracker tool usage
  • schemas/settings.schema.json:1760 — task tracker tool setting

External web sources

Search/trend context

  • DuckDuckGo search used for broader trend scan:
  • query: March 2026 Gemini CLI Codex CLI blog release notes tool calling agents terminal
  • query: March 2026 AI agent observability tracing telemetry terminal agents
  • Trend takeaway: market chatter keeps clustering around terminal agents and observability, which strengthened the decision to avoid another narration/forensics/backpressure angle and instead focus on operational state.