Versioned shared memory, atomic state transitions, conflict handling, effect logs, compensation, and controlled rollback across agents and machines.
TOOL → COMPUTER → EXECUTION FABRIC
Claude Code
From Tool to Computer
Claude Code is retracing computer history: I/O, programs, interrupts, processes, memory, scheduling, supervision, and recovery. First the tool becomes the computer. Then the computer becomes invisible infrastructure.
Interactive architecture map
The machine that is booting
Play the timeline to power each subsystem. Select a module to see what it solved and what remains unfinished.
System state Documented capability Available · partial/preview ClaudeWorld inferred gap
Playback marks when a capability entered the source-backed history. Every module card lists today’s inventory; colors, maturity, and gap statements are ClaudeWorld’s architectural analysis. Official links support current capabilities only.
14 / 15 NOW
Agent OS gaps Unpopulated expansion slots
The missing slots reveal what still prevents the parts from behaving like one reliable computer.
ClaudeWorld analysis
Unpopulated expansion slots
Three structural gaps this map focuses on—not an exhaustive list or a product roadmap.A governor for model, token, time, money, priority, fairness, cancellation, and backpressure across all work.
Always-on services that detect failure, recover, retry, verify, upgrade, and stop under explicit human policy.
Gap and maturity labels are ClaudeWorld’s architectural analysis of the documented system—not an Anthropic product roadmap.
Open the complete module responsibility ledger
CPU-00 ● Documented capability
CPU · execution core
Agent loop
- What it completed
- Reason, choose an action, observe the result, and continue until the task is done.
- ClaudeWorld inferred gap
- No unified resource governor yet chooses compute by cost, latency, and reliability across workloads.
BUS-01 ● Documented capability
I/O bus · peripherals
External world
- What it completed
- Tools and data became discoverable through a standard connection boundary.
- ClaudeWorld inferred gap
- Trust, failure, delivery, and permission semantics still vary by connector and device.
PRG-02 ● Documented capability
Programs · instruction memory
Reusable capability
- What it completed
- Working methods can be saved, discovered, invoked, and reused instead of rewritten as prompts.
- ClaudeWorld inferred gap
- Dependencies, runtime contracts, and reproducible environments are not yet one complete program model.
IRQ-03 ● Documented capability
Interrupt controller
Lifecycle events
- What it completed
- Checks and reactions can fire deterministically at specific lifecycle events.
- ClaudeWorld inferred gap
- There is no unified priority, policy, and audit plane for every event and side effect.
PKG-04 ● Documented capability
Package system
Software ecosystem
- What it completed
- Capabilities became installable, versioned bundles that teams can distribute.
- ClaudeWorld inferred gap
- Plugins already support versions, SHAs, and dependencies; platform-wide reproducibility, signed provenance, compatibility guarantees, and supply-chain policy remain incomplete.
PROC-05 ● Documented capability
Processes · multitasking
Agent processes
- What it completed
- Independent contexts can explore and work concurrently without blocking the main conversation.
- ClaudeWorld inferred gap
- Sessions and background agents expose IDs and stop controls, but there is no unified lifecycle contract for identity, signals, exit semantics, health checks, and restart policy.
SCH-06 ◐ Available · partial/preview
Scheduler · IPC
Coordination layer
- What it completed
- Workers can share tasks, message peers, run on a cadence, and execute large graphs.
- ClaudeWorld inferred gap
- Global priority, quotas, fairness, preemption, cancellation, retry, and backpressure are not one scheduler.
MEM-07 ◐ Available · partial/preview
RAM · persistent storage
Working memory
- What it completed
- The system can manage working context and retain useful local knowledge across sessions.
- ClaudeWorld inferred gap
- Auto Memory is machine-local; agents still lack coherent, versioned, transactional shared memory across devices.
SEC-08 ◐ Available · partial/preview
Kernel boundary · filesystem
Isolation & policy
- What it completed
- Actions can be constrained, shell processes isolated, and parallel file changes separated.
- ClaudeWorld inferred gap
- Built-in tools have permission controls and Bash has OS sandboxing, but isolation mechanisms differ across surfaces; there is no one end-to-end boundary for every GUI, MCP call, and secret.
CTL-09 ◐ Available · partial/preview
Task manager · observability
Control plane
- What it completed
- Humans can supervise sessions, set completion conditions, and inspect operational signals.
- ClaudeWorld inferred gap
- Telemetry exposes operational and cost metrics, but causal traces, full state diffs, deterministic replay, and outcome evidence are not yet one debugger.
REC-10 ◐ Available · partial/preview
Recovery · transactions
Restore loop
- What it completed
- Edit-tool changes and session work can be restored or resumed after a wrong turn.
- ClaudeWorld inferred gap
- Checkpoints do not track Bash or external side effects; reliable cross-agent transactions still need effect logs, idempotency or deduplication, compensation, and controlled rollback.
THE SECOND HORIZON · POST-COMPUTER
First become the computer. Then make the computer disappear.
Completion is only the bootstrap phase. Once identity, authority, state, recovery, provenance, and audit are enforceable, an agent-native system can stop making one-machine abstractions the primary user contract and organize work around intent, events, and verified outcomes.
“Disappear” does not remove hardware. Files, apps, windows, shells, and machine location stop being the primary model—but only with enforceable identity, authority, state, recovery, provenance, and audit.
The computer becomes substrate, not interface.
Apps Capabilities
Ask for a typed ability with explicit authority instead of opening and driving an application.
Files Semantic state
Use versioned entities, relationships, events, and provenance as the world model—not a folder tree.
Processes Goal actors
Schedule goals, budgets, messages, and completion contracts instead of PIDs and foreground windows.
Commands Outcome contracts
Verify that the world reached the intended state instead of treating a zero exit code as success.
Machines Execution fabric
When policy permits, place work across local, cloud, browser, edge, and embodied devices without exposing location to the goal.
MP4 · 1280×720 · H.264
SHAREABLE FILM · 16:9
Take the map with you
A silent, captioned MP4 tells the same boot sequence and ends on the missing system layers. Download it or share the file on supported devices.
Open video transcript
- START Execution core
The agent loop can reason and act, but it begins as an isolated CPU with no system around it.
- 2024-11-25 I/O bus
MCP connects the execution core to external tools and data.
- 2025-03-05 Programs
Commands turn recurring prompts into invokable programs.
- 2025-06-30 Interrupts
Hooks add deterministic lifecycle signals and guardrails.
- 2025-07-24 Processes
Subagents create isolated workers with their own context.
- 2025-10-09 Package system
Plugins make capabilities installable and distributable.
- 2025-10-16 Application layer
Skills make knowledge and workflows discoverable and reusable.
- 2025-12-05 Multitasking
Background agents keep work running without blocking the foreground.
- 2026-01-07 → 2026-01-14 Forked context + on-demand tool loading
Forked Skills and lazy tool schemas control where context is loaded.
- 2026-02-05 Multiprocess coordination + IPC
Agent Teams add shared tasks and peer messaging, still experimentally.
- 2026-02-19 → 2026-02-25 Isolated working trees + persistent local memory
Worktrees isolate file state; Auto Memory carries local knowledge forward.
- 2026-05-11 Task manager
Agent View and goals make persistent work visible and supervisable.
- 2026-05-28 Job scheduler
Dynamic Workflows coordinate large, repeatable execution graphs.
- NOW Agent OS gaps
The missing slots reveal what still prevents the parts from behaving like one reliable computer.
- BEYOND Agent-native fabric
The organizing primitives shift from machines and apps to goals, capabilities, semantic state, and outcome contracts.
It is no longer just a tool. It is not yet a complete computer.
SOURCE LEDGER · 2024—2026 The source-backed historical ledger The architecture map is the thesis. This timeline is the evidence: twelve dated milestones, each tied to an official announcement, changelog, release, or documentation page.
Verified through Claude Code v2.1.215
01 Connect
Model Context Protocol
MCP gave Claude a standard way to connect to external tools, data sources, and actions. It is a protocol boundary, not a prompt library.
Source Announcement
{
"mcpServers": {
"my-api": {
"type": "http",
"url": "https://example.com/mcp"
}
}
} 02 Package
Custom Slash Commands
Markdown files in .claude/commands/ turned recurring prompts into user-invoked workflows. The format remains compatible, but Skills are now preferred.
Source Changelog
---
description: Review the current changes
allowed-tools: Read, Grep, Glob, Bash(git diff *)
---
Review the current diff and report only actionable findings. 03 Automate
Hooks
Hooks first made lifecycle events run shell commands automatically. Today handlers can also call HTTP endpoints, MCP tools, prompts, and agents. The trigger is deterministic; prompt and agent results remain model-dependent.
{
"hooks": {
"PostToolUse": [{
"matcher": "Edit|Write",
"hooks": [{
"type": "command",
"command": "pnpm lint"
}]
}]
}
} 04 Delegate
Custom Subagents
Focused workers received their own context, tools, and permissions, then returned a summary to the calling session. Isolation reduces main-thread noise, not total token use.
Source Changelog
---
name: code-reviewer
description: Reviews code changes for correctness
tools: Read, Grep, Glob, Bash
model: sonnet
---
Review the diff, verify each finding, and return a concise report. 05 Distribute
Plugins & Marketplaces
Plugins launched as versioned bundles for commands, subagents, hooks, and MCP servers. Today they can also package Skills and LSP integrations for installation and sharing.
Source Announcement
{
"name": "team-toolkit",
"version": "1.0.0",
"description": "Shared engineering workflows"
} 06 Reuse
Agent Skills
Skills introduced reusable instructions, knowledge, scripts, and resources that Claude can discover by relevance or users can invoke directly. Optional forked execution arrived later in v2.1.0.
07 Parallelize
Background Agents
Delegated work no longer had to block the main loop. Agents could continue in the background while the primary conversation moved on to other work.
Source Changelog
08 Simplify
Forked Skills, One Model, Lazy MCP Tools
v2.1.0 added optional forked Skill execution, v2.1.3 unified Commands and Skills, and v2.1.7 made Tool Search defer MCP definitions by default. Schemas and results from tools actually used still enter context.
---
name: code-review
description: Review code changes when correctness matters
context: fork
agent: code-reviewer
---
Inspect the diff, run focused checks, and report verified findings. About 20 months of evolution · Nov 2024 → Jul 2026
Choose by control model
The right feature depends on who coordinates the work, what must stay isolated, and how much context you are willing to spend.
| Capability | Best for | Context & cost | Choose it when |
|---|---|---|---|
| MCP | External tools and data | Definitions deferred · used schemas and results enter context | Claude must read or act in another system |
| Skills | Reusable knowledge and workflows | Description at startup · full body on invoke · optional fork | A process should be repeatable and shareable |
| Subagents | Focused delegation | Separate context + returned summary | Raw exploration would pollute the main thread |
| Agent Teams | Coordinated parallel workers | High · one full context per teammate | Workers need a shared task list and direct messaging |
| Agent View | Human-supervised sessions | One full context per session | Independent jobs need one operational dashboard |
| Hooks | Lifecycle-triggered guardrails | Commands can avoid the model · prompt and agent handlers use tokens | A check or reaction must fire on a specific event |
| Schedules & routines | Recurring agent work | Variable · each run starts or continues agent work | A session or cloud task must run on a cadence |
| Dynamic Workflows | Codified large-scale orchestration | High and variable · tens to hundreds of agents | A repeatable execution graph must coordinate many agents |
Context cost describes the architecture, not a fixed token count. Actual use depends on loaded content, tool output, model, and number of active workers.
Choose the right parallel model
Compare Subagents, Agent Teams, Agent View, and worktree isolation before adding more workers.