Subagents, Agent Teams, and Durable Agents

“How are your agents different from Claude’s subagents and agent teams?” is the first question every time we demo. The short answer is lifecycle: Claude’s primitives orchestrate inside a single session. Ours form a durable engineering org that remembers what it learned yesterday.

What are Claude subagents and agent teams?

Session-scoped orchestration primitives inside Claude Code. Subagents are the contractor pattern: the parent dispatches a worker and waits on it, context stays isolated, and communication is hub-and-spoke through the orchestrator. Agent teams (the newer experimental feature behind CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) upgrade that to peer-to-peer: full Claude sessions with their own context windows, a shared task list, and direct messaging between teammates.

If you already run Claude Code for parallel work, use agent teams over plain subagents; the coordination is better. But in both shapes the lifecycle is identical: the team is reconstituted from scratch each invocation, runs, and dies when the session ends.

How is a durable agent different from a spawned one?

It exists across sessions and keeps what it learned. In our setup, the CTO, Staff Engineers, Engineers, and QA Engineers are durable identities. They persist across sessions and across projects, with their own configured memory providers: Hindsight for the CTO, Holographic for the engineers doing the building, different stacks for different roles. A Claude agent team is reconstituted from scratch each invocation; our engineers show up tomorrow remembering what happened today.

They’re equipped with harnesses where memory compounds. They store knowledge graphs of the system, of you, and of how they need to operate as an identity to get better in future cycles. They change and evolve, and you feel it when you work with them day to day.

Why does a durable agent beat spinning one up for each session?

Because a spawned agent starts from zero every time, and a durable one starts from everything it has learned. With Claude subagents or teams, context stays isolated and dies when the session ends. Every lesson from the run is gone, and next time you do it all over again. The agent isn’t growing with you. Try to get it to remember a lesson and you’ll see the same mistake a few days later, because there is nowhere for the lesson to live. Agent teams’ “shared task list” doesn’t change that: it’s a coordination device for one run, not institutional memory. Our agents write home after every ticket: what they learned about the codebase, what surprised them, which anti-patterns to avoid next time. By the next session, that knowledge is in their context before they read the ticket. That loop is what turns a model into a colleague. A spawned worker is as good on its hundredth ticket as on its first. A durable agent on its hundredth ticket carries ninety-nine tickets of your schemas, your conventions and your review bar, and it gets there without you repeating yourself.

Claude’s primitives have no equivalent to this.

What does a durable agent look like after 100 tickets?

Specialised. Our frontend engineer has completed over 100 tickets of frontend work and has become specialised through what it learned: the schemas it has memorised, the TDZ patterns it has learned to avoid, the review bar it has calibrated against. A Claude subagent is a fresh session, and you can’t guarantee any particular result.

With our agents you become more and more hands-off, because they learn you, the system, and their role. With Claude subagents you’re still the QA, still the sceptic validating everything that comes through. We say that from running both.

It’s an incredible feeling, you notice the difference once you start using it, and it’s hard to go back.

Do durable agents run on a different model?

No. Our LLM provider is Claude. We lean on Opus 4.7 Max to get the best output out of every agent. The difference isn’t the model; it’s the architecture on top of it. Claude’s primitives are the right answer when you want to parallelise a task inside a session. Durable agents are the right answer when you want an engineering org that compounds.