Brain, Hands, Session
The architectural choice that makes Managed Agents interesting isn't the API surface. It's the split.
In the old design, an agent's reasoning loop and its execution sandbox lived in the same process. To start thinking about a problem, the agent first had to wait for a container to boot. Anthropic's engineering blog describes the new approach using an operating-system-style virtualization pattern: the run is split into three distinct components.
🧠 Session
An append-only log of every event — the durable record that survives crashes and enables replay.
⚙️ Harness
The loop that calls Claude and routes its tool calls — the "brain" of the agent runtime.
📦 Sandbox
The execution environment — the "hands" that actually run code the model writes.
"Managed Agents is a meta-harness in the same spirit, unopinionated about the specific harness that Claude will need in the future. Rather, it is a system with general interfaces that allow many different harnesses." [1]
The numbers Anthropic published are worth pausing on. After the decoupling, p50 time-to-first-token dropped by roughly 60%, and p95 dropped by more than 90% — not because the model got faster, but because sessions stopped waiting on containers to spin up before inference could begin. [1]
p50 is what the demo feels like. p95 is what your on-call engineer feels at 3 a.m. when an agent stalls on a long task and the user has already alt-tabbed away.
There is also a quieter benefit. Anthropic's launch post claims the platform's optional Outcomes feature — currently in research preview, not the standard public beta — improves task success "by up to 10 points over a standard prompting loop, with the largest gains on the hardest problems." [12] That is not the kind of number you can replicate with a weekend of YAML, but it's also not a free win in the public beta — you have to apply for access. [2]
Who's Actually Winning the Orchestration Layer
Here is where the optimism needs cold water.
VentureBeat's directional research surveyed 56 organizations with 100+ employees in January 2026 and 70 in February. The results paint a clear picture of the current landscape:
That January-to-February jump for Anthropic — from a flat 0% to 5.7% — is the most interesting datapoint in the whole survey. And it predates the Managed Agents launch; the growth came from Claude's existing tool-use API, not the new hosted runtime. The trajectory is real. The base is small.
Microsoft's Strategy
Winning enterprise orchestration the way it has won every other enterprise market: through licensing relationships, predictable capacity-based pricing (Copilot Studio capacity packs include 25,000 Copilot Credits at $200/pack/month), and a willingness to be model-agnostic. [5][15]
Anthropic's Strategy
Going the other way: bet hard on a single model, optimize the harness around it, and price it like compute. That pricing model is genuinely novel and genuinely worth scrutinizing.
The Pay-Per-Second Math
Managed Agents charges $0.08 per active session-hour on top of standard Claude API token rates, with web searches billed separately at the standard $10 per 1,000 searches. [12][13] Idle time — while the session is waiting on input or tool confirmations — does not count toward runtime billing. [13] That sounds wonderful, until you do the arithmetic.
$0.08
Per Active Session-Hour
On top of standard Claude token rates. Idle time is free.
$58
Single Always-On Agent/Mo
730 hours × $0.08 ≈ $58.40 in pure runtime overhead — before a single inference token.
$461
24-Agent System/Mo
24 agents × 8h/day × 30 days × $0.08 ≈ $460.80, before any inference costs.
These are illustrative calculations from Anthropic's published runtime rate, not figures Anthropic has published. Scale to a real enterprise — a few hundred agents, all genuinely useful — and runtime starts to rival the model bill.
The case Anthropic is making is that the capability unlocked is so much higher that the math works anyway. The General Legal CTO they cite in the launch post puts it neatly: "Now, with Managed Agents it can code up any tool it needs on the fly, allowing it to handle virtually any user query." [12]
That's the dream — an agent that authors its own tools at runtime, runs them in a sandbox you didn't have to build, and leaves a replayable session log when it's done. Anthropic's launch post names Notion, Rakuten, Asana, Vibecode, Sentry, Atlassian, General Legal, and Blockit as early customers. [12]
About the Author
David Proctor
VP of AI at Trilogy. Writing about AI infrastructure, agent protocols, and what actually works in production.
What This Newsletter Covers
- AI infrastructure deep dives
- Agent protocols and orchestration
- What actually works in production — not just what demos well
- Honest analysis of vendor trade-offs
If you found this analysis useful, the best thing you can do is share it with the engineer on your team who's still hand-rolling agent infrastructure.