Agent Atlas / Coding agents

google-gemini/gemini-cli

A terminal-first coding agent whose public source exposes tool scheduling, hooks, policy, context assembly, extensions, and specialized workers.

Coding agent CLI Full dossier Reported + sourced Reviewed August 26, 2026

Intended operator

Developers operating an agent in a terminal or through an integrating client.

Primary surfaces

terminal client / built-in tools / MCP and extensions

Tracked repository

google-gemini/gemini-cli

Teaching mechanism

Hooks can add context or stop work before execution

BeforeAgent and related hook boundaries let the runtime attach delimited context or block a run before the next agent step.

Evidence limitA documented hook boundary does not establish that every extension is safe or that every client exposes the added context.

Inspect the source

System Map

Eight inspection boundaries

Model boundary

The CLI assembles provider requests and tool declarations around the selected Gemini or compatible model path.

Context and memory

Workspace files, conversation state, settings, hooks, and extension context can shape the model-visible request.

Tools and execution

Built-in file, shell, web, and MCP tools pass through scheduler and policy code.

Permissions and isolation

Approval modes, policy checks, trusted folders, and sandbox options separate a proposal from an allowed effect.

Sessions and recovery

Session and tool lifecycle state support resumption, cancellation, and visible failure handling.

Delegation

Specialized subagents receive focused prompts, tool scopes, and independent context before returning results.

Human control

Terminal status, approval, command, and issue-maintenance surfaces keep operator decisions in the loop.

Evidence surface

Public commits, tests, tool events, scheduler behavior, and issue-maintenance receipts expose runtime pressure.

Learn Through This Project

8 lessons

Chapter 1 Before the First Model Request Chapter 2 The Invisible Prompt Chapter 3 Tool Calls: Contract, Authorization, Execution Chapter 4 Loops, Workflows, and Recovery Chapter 5 Delegation Is a Distributed System Chapter 7 The Approval Trap Chapter 8 The Human Control Loop Chapter 9 Trust Must Make Its Case

Compare Its Boundaries

8 labs

Comparison labTool Contracts Comparison labGoverned Memory Comparison labDelegation Receipts Comparison labApproval Boundaries Comparison labHuman Workbench Comparison labSessions And Recovery Comparison labEvidence And Evaluation Comparison labMaturity Signals

Latest Reporting

8 articles

Before Codex Trusts a Workspace, It Stops Letting the Workspace Choose Its Helpers

September 3, 2026 — A merged Codex change moves automatic pre-trust helper lookups to known system locations, sanitizes plugin-sync Git, and makes diagnostics inspect a workspace-influenced path without running what they find. openai/codex, google-gemini/gemini-cli

A Restricted Workspace Should Not Start Its Own Tools

August 29, 2026 — Gemini CLI now gives an explicitly restricted or untrusted workspace a narrower configuration path: repository-defined tools, MCP servers, policies and telemetry no longer accompany its A2A server into startup. google-gemini/gemini-cli

Coding Agents Are Building Themselves—But Git Can Prove Only Part of It

August 11, 2026 — Across 40,345 unique commits in ten public agent codebases, coding-agent credit becomes sharply more visible in 2026—and the missing receipts matter as much as the names that appear. openai/codex, google-gemini/gemini-cli, anomalyco/opencode, QwenLM/qwen-code, cline/cline, aaif-goose/goose, OpenHands/OpenHands, charmbracelet/crush, mistralai/mistral-vibe, Kilo-Org/kilocode

The Supervision Derby finish: which coding-agent repo made control more real?

July 31, 2026 — The finish is a split decision: Codex makes filesystem policy travel into capability discovery; Qwen Code makes a fork’s tool gate survive resume; OpenHands and Gemini CLI make narrower but useful controls reachable. openai/codex, google-gemini/gemini-cli, QwenLM/qwen-code, OpenHands/OpenHands

The Harness Was Taking the Test Too

July 30, 2026 — OpenAI says two context choices lifted GPT-5.6 Sol from 13.3% to 38.3% on the ARC-AGI-3 public set. Codex source shows those “settings” are really a production memory system—and rival harnesses reveal how much of it can be copied. openai/codex, pydantic/pydantic-ai, earendil-works/pi, NousResearch/hermes-agent, mistralai/mistral-vibe, anomalyco/opencode, QwenLM/qwen-code, google-gemini/gemini-cli, cline/cline, aaif-goose/goose, OpenHands/OpenHands

The Supervision Derby: four coding-agent control surfaces enter the first Diff Derby

July 29, 2026 — The opening call gives Codex a narrow provisional edge for making subagent instruction precedence survive delegation and resume—but Qwen Code, Gemini CLI, and OpenHands each put a different kind of operator control on the track. openai/codex, google-gemini/gemini-cli, QwenLM/qwen-code, OpenHands/OpenHands

An Agent’s Answer Begins in the Harness

July 26, 2026 — The model only sees the task after a runtime has edited its context, carried forward state, and decided what survived the last turn—so the first question after a bad run is what the harness made it see. QwenLM/qwen-code, cline/cline, block/goose, google-gemini/gemini-cli

Before You Blame the Model, Inspect What the Harness Made It See

July 26, 2026 — A model's next move is shaped by the tool output, trust state, history, and recovery rules a harness hands it—so a bad run is not automatically a bad model. QwenLM/qwen-code, cline/cline, aaif-goose/goose, google-gemini/gemini-cli

Primary Sources

Google Gemini CLI repository

Primary public implementation and documentation source.

Open Questions

Can every injected context and policy result be reconstructed from the operator-facing session record?

Evidence state describes what this Atlas profile can support today. It is not a product score and it is not evidence that uninspected capabilities are absent.