The agent-runtime landscape, mid-2026

Before committing agentd to its final shape, the lab ran a deliberately oversized research pass: 107 agents sweeping 25 primary sources, extracting 125 claims, each adversarially verified before entering the corpus. The survey that resulted is the lab's map of the mid-2026 landscape.

Its taxonomy sorts the ecosystem by where the loop lives:

• Framework-embedded (LangGraph-style): the loop is a library in your process; maximal flexibility, minimal operational contract.
• Durable-workflow engines (Temporal-adjacent): the loop is replayable history; strong guarantees, heavyweight embrace.
• Sandboxed runtimes: the loop runs inside enforced isolation; the contract is the boundary itself.
• Managed platforms: the loop is someone else's service; operational ease traded against inspectability.

Three findings shaped agentd directly. First, supervision is the scarce property — most stacks can run an agent, few can contain one, and containment is what production actually asks for. Second, the tool interface converged: MCP won the protocol argument, so a runtime gains nothing by inventing tool plumbing. Third, cold start and footprint compound — a fleet of long-tail, mostly-idle agents is economically shaped by per-agent overhead, which is the argument for a 3 MiB binary with a sub-millisecond start.

The survey is also the provenance story for the runtime's design docs: when a claim in the vision document cites the ecosystem, it cites this corpus — with the verification trail attached.