There’s a moment in any multi-agent project when you realize you’ve been managing agents the hard way. You’re running a research agent in one terminal window, a code reviewer in another, and a documentation agent in a third. Each has its own PID, its own context, its own scrollback. Switching between them feels less like orchestrating agents and more like playing Tetris with terminal windows.
Claude Code shipped something in May 2026 that changes the shape of this problem: the claude agents command, which opens an agent view — a single dashboard showing every background session, their status, and what needs your input. It’s a CLI-level primitive that turns Claude Code from a single-agent tool into something closer to an orchestration layer.
What claude agents Actually Does
The command itself is deceptively simple: claude agents in any terminal window opens the agent view. That’s the visible part. The interesting part is what it’s built on top of.
The May 2026 Claude Code release added a set of new flags to the claude agents command that make it genuinely useful for multi-agent work: --cwd <path> filters to show only sessions started under a specific directory, --add-dir adds directories to the session scope, --settings and --mcp-config let you control the environment per session, --plugin-dir manages plugins, --permission-mode controls permission levels, --model selects the model, --effort sets the effort level, and --dangerously-skip-... handles safety bypasses. The breadth of these flags tells you the design intent: this isn’t a toy dashboard, it’s a full environment control surface for background agent sessions.
Background sessions themselves are not new — Claude Code has had background execution for a while. What changed is the observability layer. Before agent view, a background session running in the background meant you either polled it periodically or waited for a notification. Now you have a live view of every session, their current state, and what they’re blocked on.
The Parallel Session Problem
Aniket has been running multi-agent pipelines in various forms for over two years. The hardest part has never been making a single agent do useful work — it’s been the coordination problem: how do you run Agent A and Agent B against the same codebase without them stepping on each other? How do you know which agent is doing what? How do you handle the case where Agent A is blocked waiting for a file that Agent B is still writing?
The claude agents --cwd <path> flag addresses a specific version of this problem: if you’re working in a monorepo with multiple projects, you can scope agent sessions to specific directories and filter your dashboard to show only the sessions relevant to the work you’re doing. This is a real quality-of-life improvement for anyone running Claude Code across a multi-project workspace.
The claude agents flags for --model and --effort go further. In a multi-agent pipeline, different agents have different performance profiles. A research agent might need Sonnet 4.5’s speed for broad exploration but switch to Opus 4.7 for synthesis. A code reviewer might run on a fast model for routine changes but escalate to Opus 4.7 for security-sensitive diffs. The ability to set model and effort per session — from the same orchestration interface — is where the agent view stops being a monitoring tool and starts being a control plane.
Background Sessions and Daemon Reliability
The May 2026 changelog also mentions improvements to daemon reliability and background session handling. This is the unsexy part that matters enormously in practice.
Background agents fail in ways that foreground agents don’t. A process gets killed when you close the terminal. A long-running agent gets silently dropped when the terminal session times out. The daemon that manages background sessions crashes and takes its running agents with it. These aren’t hypothetical failure modes — they’re the reason most people don’t trust background agents with important work.
The daemon improvements in the May release make background sessions meaningfully more reliable. When combined with the agent view’s monitoring capabilities, you now have both the control plane and the reliability layer needed to trust background agents with real work.
What This Means for Orchestration
The pattern that’s emerging from these Claude Code updates is clear: the CLI is becoming an orchestration interface, not just an execution interface. You don’t open Claude Code to run a single task — you open it to run a system of agents, monitor their state, and intervene when they need input.
Aniket’s ACO System uses a pipeline architecture — PM → Planner → Architect → Developer → QA → Human Review. That’s six stages, some of which can run in parallel for independent tasks. Running that pipeline requires knowing which stage each task is in, which agents are blocked, and which need human input. The agent view is the closest thing to a terminal-native control plane for exactly that problem.
The claude agents command isn’t a replacement for dedicated orchestration frameworks. But for engineers running agent workloads from a terminal — which is still most of the people building this stuff — it’s the difference between managing agents with ps and managing them with a proper monitoring system.
The Fast Mode Default Shift
One more detail from the May changelog worth noting: fast mode in Claude Code now defaults to Opus 4.7. This matters for the orchestration story because fast mode is what background sessions use when you don’t explicitly specify a model. The implication is that every background session started without an explicit model flag now runs on Opus 4.7 — the same model that scores 87.6% on SWE-bench Verified.
This is a meaningful default. Fast mode has always been about balancing speed and capability. Setting Opus 4.7 as the default means the baseline quality for background sessions is now significantly higher than it was a month ago. If you’ve been running background agents on older defaults, the upgrade is free and the capability improvement is real.
The agent view, the new CLI flags, the daemon reliability improvements, and the Opus 4.7 fast mode default are all separate changes. Together they describe a coherent direction: Claude Code is becoming the terminal-native layer for multi-agent orchestration. Not because it was rebranded or given a fancy new UI, but because the CLI primitives needed to run, monitor, and control multiple agents in parallel are now in place.