AGENT TOOL LANDSCAPE · CURRENT SNAPSHOT

Three routes toward
the agent computer.

Claude Code, Codex, and Grok Build are absorbing the old computer stack from different directions. This map compares documented capability, current emphasis, inferred momentum, and what none of them has closed yet.

Evidence boundary

OFFICIALSolid status markers mean a capability is documented in current first-party material. Limited means preview, experimental, surface-bound, or incomplete.

INFERENCEDashed trajectory lines are ClaudeWorld analysis. They are not vendor roadmaps, promises, benchmarks, or product scores.

ONE COMPUTER MODEL · THREE TRAJECTORIES

The same 11 system layers, viewed through three products

Select a product to isolate its trajectory. Compare all keeps every marker visible without inventing a winner.

Current trajectory

Claude Code

Moving from an interactive coding tool toward an Agent OS and collaboration control plane around sessions, skills, hooks, teams, goals, and human supervision.

Included surfaces
CLI · Desktop · Web · IDE · SDK
SCOPE NOTE
Claude Code as a CLI, desktop/web/IDE surface, and Agent SDK—not the general Claude chat product.
Current emphasis
Lifecycle controlReusable capabilitiesAgent coordinationHuman-in-the-loop operations
Direction vectors

Agent collaboration control plane ClaudeWorld inference · Strong signal

Agent View, teams, workflows, scheduled work, and goals shift the interface from one chat toward operating many workers.

Skills as agent programs ClaudeWorld inference · Strong signal

Skills, plugins, hooks, and MCP increasingly package repeatable work rather than one-off prompts.
Still unresolved for this product · ClaudeWorld inference
  • One resource scheduler across every session and surface
  • Transactional shared state across devices
  • Effect-safe recovery beyond edit checkpoints
Current trajectory

Codex

Building a local-to-cloud execution fabric: parallel long-running workers, isolated worktrees, skills, automations, review, and remote supervision.

Included surfaces
CLI · App · IDE · Cloud · SDK
SCOPE NOTE
Codex across CLI, IDE, desktop app, cloud, and SDK. Availability differs by surface.
Current emphasis
Parallel long-running workLocal ↔ cloud continuityReview and approvalsSkills and plugins
Direction vectors

Agent command center ClaudeWorld inference · Strong signal

The app explicitly reframes development from pairing with one agent to supervising coordinated teams across long tasks.

Computer as an execution host ClaudeWorld inference · Strong signal

Goals, skills, plugins, schedules, cloud environments, and remote control move the user model above files and terminals.
Still unresolved for this product · ClaudeWorld inference
  • Canonical semantic state across local and cloud
  • Uniform authority across shell, MCP, browser, and computer use
  • Transactional rollback for real-world side effects
Current trajectory

Grok Build

A compatibility-first, portable local agent runtime that absorbs existing Claude, AGENTS.md, MCP, ACP, plugin, hook, and custom-model conventions.

Included surfaces
CLI · TUI · Headless · ACP
SCOPE NOTE
Grok Build, the coding-agent runtime. This does not merge in the separate Grok chat or hosted research multi-agent product.
Current emphasis
InteroperabilityLocal operator controlExtensible runtimeEnterprise endpoint policy
Direction vectors

Portable agent shell ClaudeWorld inference · Strong signal

ACP, custom models, MCP, and compatibility layers position the harness as a replaceable runtime around many ecosystems.

Governed local control plane ClaudeWorld inference · Strong signal

Sessions, dashboard, permissions, sandbox, monitoring, and enterprise policy converge on managed execution at the endpoint.
Still unresolved for this product · ClaudeWorld inference
  • Always-on supervisor that survives host loss
  • Cross-host scheduling and resource placement
  • Signed packages and effect-safe transactional recovery
Documented Documented · limited No official evidence found Not assessed

“No official evidence found” means no evidence was found in the reviewed sources as of 2026-07-19; it does not prove a feature is absent.

Editorial emphasis · CW Emphasis badges are ClaudeWorld inference, not official maturity labels or product scores.

Included surfaces Coverage means at least one included surface has first-party evidence. It does not imply feature parity across every surface.

THE SHARED EMPTY SPACE · ClaudeWorld inference

What none of the three has turned into one coherent system

All three still execute through files, shells, processes, Git, operating-system sandboxes, browsers, and SaaS APIs. The post-PC break happens only when these missing layers become native.

STATE / 01 · CW

Coherent semantic state

Versioned, conflict-aware state shared across agents, devices, and external systems—not just transcripts and files.

KERNEL / 02 · CW

Global resource kernel

One place for priority, budgets, quotas, placement, cancellation, retries, and backpressure.

AUTH / 03 · CW

Uniform capability authority

Task-scoped identity and permissions across shell, files, browser, MCP, GUI, secrets, and hosted tools.

TRACE / 04 · CW

Causal effect ledger

A trace that joins decisions, tool calls, state diffs, costs, approvals, and real-world outcomes.

TXN / 05 · CW

Transactional recovery

Idempotency, deduplication, compensation, and controlled rollback for external side effects.

FABRIC / 06 · CW

Location-transparent execution

Goals move safely across local, cloud, browser, edge, and devices without inheriting one computer’s limits.

Official evidence ledger Every coverage marker links to first-party documentation, announcements, or repositories reviewed for this dated snapshot.

Claude Code

Codex

Grok Build

TOOL → COMPUTER → EXECUTION FABRIC

The destination is not a better terminal.

It is a goal runtime that can place work, control authority, preserve state, prove outcomes, and recover across machines without exposing the old computer as the unit of work.

See the evolution map Choose a parallel model Inspect Open Systems