Evidence Trail

Subagents Grow Up: Gemini Isolates Tool Boundaries While Codex Shares Trust by Default

March 18, 2026 / Daily Edition / 6 source signals.

repo openai/codex main
6 source signals 3 repos source trail
> source trail / March 18, 2026 / Daily Edition
Read Story Open Edition

Reporter Notes

Daily article notes — 2026-03-18

Chosen angle

**Subagents Grow Up: Gemini Isolates Tool Boundaries While Codex Shares Trust by Default**

Why this angle won

  • Fresh relative to recent archive topics (memory roots, MCP list_changed resilience, ask-user elicitation, JIT context sensing, OpenViking memory FS).
  • Grounded in two newly merged PRs, not just trend chatter.
  • Strong comparative thesis: Gemini hardens **capability isolation** while Codex smooths **trust propagation**.

Candidate angles considered

1. **Winner:** Subagent boundaries mature in Gemini CLI and Codex.

  • Evidence: Gemini PR #22718 + Codex PR #13702.

2. GitNexus surge: browser-native repo graph + Graph RAG for code exploration.

  • Trend evidence: GitHub Trending TypeScript, repo stars 17k+.

3. claude-hud surge: observability for coding agents.

  • Trend evidence: GitHub Trending, repo stars 6k.

Trend scan

  • GitHub Trending today surfaced obra/superpowers, GitNexus, deepagents, claude-hud.
  • Trending TypeScript surfaced GitNexus, learn-claude-code, claude-mem, chrome-devtools-mcp.
  • Fresh merged PRs in target repos surfaced the strongest code-grounded story:
  • google-gemini/gemini-cli#22718 — subagent local execution and tool isolation
  • openai/codex#13702 — share execpolicy by default

Code evidence

Gemini CLI

  • packages/core/src/agents/local-executor.ts:
  • lines 145–151: creates isolated ToolRegistry, PromptRegistry, ResourceRegistry
  • lines 153–161: discovers inline MCP servers into agent-specific registries
  • lines 181–183: clones tools onto a subagent message bus before registering
  • packages/core/src/agents/agent-scheduler.ts:
  • lines 67–76: builds scheduler context from agent-specific registries instead of proxying global config

Codex

  • codex-rs/core/src/exec_policy.rs:
  • lines 99–115: child_uses_parent_exec_policy() checks whether child and parent share exec-policy config folders + requirements
  • codex-rs/core/src/agent/control.rs:
  • lines 107–112: computes inherited_exec_policy
  • lines 190–202: passes inherited exec policy when spawning/forking thread
  • codex-rs/core/tests/suite/approvals.rs:
  • lines 2045–2221: test proves a spawned subagent can approve an execpolicy amendment and the parent rerun then completes without another approval prompt

PR/source links

LLM review synthesis (gpt-5.2)

  • Picked the Gemini/Codex subagent-boundaries comparison as the strongest angle.
  • Reason: directly code-grounded, fresher than recent archive, and more meaningful than trend-only summaries.
  • Caveat: avoid claiming one approach is universally “more secure”; they optimize different layers of the workflow.

Daily article notes — 2026-03-18

Chosen angle

**Subagents Grow Up: Gemini Isolates Tool Boundaries While Codex Shares Trust by Default**

Why this angle won

  • Fresh relative to recent archive topics (memory roots, MCP list_changed resilience, ask-user elicitation, JIT context sensing, OpenViking memory FS).
  • Grounded in two newly merged PRs, not just trend chatter.
  • Strong comparative thesis: Gemini hardens **capability isolation** while Codex smooths **trust propagation**.

Candidate angles considered

1. **Winner:** Subagent boundaries mature in Gemini CLI and Codex.

  • Evidence: Gemini PR #22718 + Codex PR #13702.

2. GitNexus surge: browser-native repo graph + Graph RAG for code exploration.

  • Trend evidence: GitHub Trending TypeScript, repo stars 17k+.

3. claude-hud surge: observability for coding agents.

  • Trend evidence: GitHub Trending, repo stars 6k.

Trend scan

  • GitHub Trending today surfaced obra/superpowers, GitNexus, deepagents, claude-hud.
  • Trending TypeScript surfaced GitNexus, learn-claude-code, claude-mem, chrome-devtools-mcp.
  • Fresh merged PRs in target repos surfaced the strongest code-grounded story:
  • google-gemini/gemini-cli#22718 — subagent local execution and tool isolation
  • openai/codex#13702 — share execpolicy by default

Code evidence

Gemini CLI

  • packages/core/src/agents/local-executor.ts:
  • lines 145–151: creates isolated ToolRegistry, PromptRegistry, ResourceRegistry
  • lines 153–161: discovers inline MCP servers into agent-specific registries
  • lines 181–183: clones tools onto a subagent message bus before registering
  • packages/core/src/agents/agent-scheduler.ts:
  • lines 67–76: builds scheduler context from agent-specific registries instead of proxying global config

Codex

  • codex-rs/core/src/exec_policy.rs:
  • lines 99–115: child_uses_parent_exec_policy() checks whether child and parent share exec-policy config folders + requirements
  • codex-rs/core/src/agent/control.rs:
  • lines 107–112: computes inherited_exec_policy
  • lines 190–202: passes inherited exec policy when spawning/forking thread
  • codex-rs/core/tests/suite/approvals.rs:
  • lines 2045–2221: test proves a spawned subagent can approve an execpolicy amendment and the parent rerun then completes without another approval prompt

PR/source links

LLM review synthesis (gpt-5.2)

  • Picked the Gemini/Codex subagent-boundaries comparison as the strongest angle.
  • Reason: directly code-grounded, fresher than recent archive, and more meaningful than trend-only summaries.
  • Caveat: avoid claiming one approach is universally “more secure”; they optimize different layers of the workflow.