Anthropic Managed Agents vs self-hosted Claude Code

Anthropic shipped Managed Agents in late 2025 as a hosted service for production AI agents. Around the same time, Claude Code matured into a developer-grade toolkit you can run locally against your own repo and data. Both work for revenue teams. They are not the same product, and the choice between them is one of the higher-stakes decisions a CRO will make this year.

This post covers what each option is, the cost and control tradeoffs, the team shapes that fit each, and a working migration path between them. The framing comes from the GTM subagent engagements we have shipped across both architectures.

What each option is

Managed Agents is Anthropic's hosted agent service. You define the agent's tools, prompts, and capabilities through their console or API. Anthropic runs the execution loop, the retries, the scaling, the logging, and the model invocations. You see the agent as a hosted endpoint your applications call.

Self-hosted Claude Code is the local agent toolkit. A Claude Code installation in your repo, configured via CLAUDE.md, with subagents defined as folders, MCP servers connecting to your CRM and warehouse, and the orchestration loop running on your own infrastructure. Anthropic provides the model; everything else is yours.

Both use the same underlying Claude models. Both can call the same MCP servers. The difference is the surrounding system. Managed Agents wraps the loop. Self-hosted Claude Code makes you write the loop.

What revenue teams care about

Six things matter when picking between them. Cost. Control. Customization. Compliance. Operational burden. Team fit.

Cost: self-hosted is cheaper at low and medium volume. You pay for model tokens through the standard Anthropic API and nothing else. Managed Agents adds a per-execution platform fee that's reasonable per call but compounds at scale. The crossover for typical revenue work lands around 50,000-200,000 agent executions per month. Below that, self-hosted wins on cost. Above, the operational savings from Managed Agents start to dominate.

Control: self-hosted wins by a wide margin. The prompts live in your repo. The retry logic is your code. The scaling is your machines. If a model update changes behavior, you can pin and roll out on your timeline. With Managed Agents, you opt into Anthropic's defaults and Anthropic's update schedule. For most revenue use cases, the defaults are good. For sensitive ones, they may not be.

Customization: both are customizable on the layers revenue teams care about (prompts, tools, ICP definitions, voice rules). Self-hosted wins on the layers below that (orchestration, retry logic, parallelism). Most revenue teams don't need the lower layers customized. The teams that do are usually doing real-time customer-facing work, which is a different category.

Compliance: Managed Agents runs in Anthropic's environment. Data flows through their infrastructure. For most B2B revenue data, that's fine. For HIPAA or SOC 2 Type II environments with strict data residency requirements, self-hosted may be required. Anthropic's enterprise contracts cover most cases, but check the specifics against your security questionnaire before committing.

Operational burden: Managed Agents wins here by a wide margin. Self-hosted Claude Code requires someone to own the repo, run the cron, watch the logs, handle the retries, and update the dependencies. For a team of three, that's a meaningful share of one engineer's time. Managed Agents takes most of it off the table.

Team fit: self-hosted fits teams that have at least one engineer or technical RevOps lead. Managed Agents fits teams without that bench, or teams whose technical capacity is committed elsewhere. The team shape is often the deciding variable, not the cost math.

When Managed Agents fits

We recommend Managed Agents for three team profiles.

The first is large enterprises with high agent volume and a strong preference for hosted services. Tens of thousands of executions per day, multiple business units running agents, and a procurement organization that prefers SaaS contracts over engineering ownership. Managed Agents fits cleanly here.

The second is teams running customer-facing agents that need uptime guarantees. An agent that powers a chat experience, a real-time recommendation engine, or a transactional workflow benefits from Anthropic's SLA more than from self-hosted flexibility. The 99.9% target is meaningful when the agent is the product surface.

The third is teams that have no engineering capacity and are not going to hire any. A 50-person company with a strong revenue function and no in-house engineering can deploy a Managed Agents setup with a thoughtful RevOps lead, a consulting partner, and a quarterly review cadence. The same team trying to run self-hosted Claude Code without an engineer will end up with a broken repo within a quarter.

When self-hosted Claude Code fits

We recommend self-hosted Claude Code for three different profiles.

The first is technical founder-led companies that want full control over the agent surface. The CTO or technical founder reads the repo. The CLAUDE.md is treated as the brand and voice contract. Subagents get committed and reviewed like any other code. This is the modal profile for our GTM subagent engagements: a team that already runs Claude Code internally and wants the revenue function on the same stack.

The second is teams replacing AI SDR seats with a custom build. The economics of the swap (covered in replacing BDR seats with subagents) work best when the team owns the repo and can iterate weekly. Managed Agents adds platform fees per execution; self-hosted adds engineering ownership. For high-volume outbound work with a small senior team, self-hosted wins on both cost and iteration speed.

The third is teams handling sensitive data that needs to stay on their infrastructure. Healthcare provider data, financial advisor data, or any segment where data residency matters. Self-hosted Claude Code lets you keep the data layer entirely on your machines and only ship the redacted prompts to the model API.

The cost math, worked out

A realistic mid-market outbound build runs 30,000-60,000 agent executions per month across signal capture, ICP scoring, contact discovery, and drafting. At that volume, the model cost on standard Claude API runs $1,200-$3,000 per month. The platform fee on Managed Agents adds roughly $1,500-$4,000 per month depending on tier. Self-hosted runs the same model cost plus roughly $1,000-$2,500 per month in engineering time (10-25% of one engineer's bandwidth for ownership).

The numbers come out close. The deciding factor is rarely the dollar amount. It's whether the team has the engineer to spare. If yes, self-hosted; the dollars go to a person on your team and the IP stays in your repo. If no, Managed Agents; the dollars go to Anthropic and the operational burden disappears.

The migration path

Most teams we work with start self-hosted and migrate selectively. Start in Claude Code because the iteration speed is higher and the team learns the playbook faster. After 90-180 days, the team knows which subagents are stable, which need constant tuning, and which run at high volume. Promote the stable, high-volume agents to Managed Agents. Keep the tuning-heavy ones in self-hosted.

Migration in this direction is mostly mechanical. A self-hosted subagent with a CLAUDE.md, a SKILL.md, and clean tool definitions ports to Managed Agents in a few days. Migration the other way is harder because Managed Agents abstracts some of the orchestration. Going from self-hosted to Managed Agents is a graduation. Going the other way is a rebuild.

A few teams skip self-hosted and start with Managed Agents. That works if the team is large enough to justify the platform fee from day one and the agent work is well-scoped enough to define cleanly upfront. We see this profile more often in enterprise customers and almost never in Series A through C companies.

Our default recommendation

For Series A through C B2B companies with a technical founder or CTO: start self-hosted. The CLAUDE.md and subagent repo become part of your codebase. The team learns the playbook by editing it. Migration to Managed Agents is available when you outgrow the operational burden.

For PE portcos and post-take-private rebuilds: start self-hosted with a fixed-fee engagement to ship the initial build. The portfolio company gets a working repo and a trained operator. Anthropic Managed Agents is the longer-term destination if volume scales meaningfully, but the rebuild should happen in code first.

For enterprise customers or teams without engineering capacity: start with Managed Agents and a strong RevOps lead. Treat the agent definitions like any other vendor configuration: version-controlled, reviewed, and updated on a cadence. The lift is smaller and the operational risk is lower.

For the architecture that sits inside the self-hosted choice, see Claude Skills vs MCP for GTM stacks. For the working CLAUDE.md template, see how to write a CLAUDE.md for your revenue team. For the orchestration patterns that scale across either choice, see three subagent orchestration patterns. We deploy either architecture as a fixed-fee engagement and recommend the right one based on your team shape and volume.