Agent Orchestration 2026¶
Shared thesis across two Coding Agents Conference talks (zach-lloyd, niels-bantilan): 2026 is the year orchestration goes from novelty to core infrastructure. Lloyd stakes the consumer/dev-tooling side; Bantilan stakes the platform/durability side.
Lloyd's framing (Warp)¶
- Engineers are now "orchestrators of fleets they don't even run locally anymore."
- Local laptops hit physical limits with 4–5 concurrent agents against a million-line codebase → laptop-limit.
- 2026 goal: make launching an agent into the cloud as frictionless as starting a local conversation, without losing visibility or programmability.
Bantilan's framing (Union / Flyte)¶
- Durability ≠ observability ≠ debuggability; you need all three (durable-observable-debuggable-agents).
- Replay logs + checkpointed state are the substrate (replay-log).
Why this matters¶
Lloyd is the interface layer, Bantilan is the runtime layer. Together they describe the emerging orchestration stack: UI/UX on top, durable-task substrate below, sub-agents (subagent-architecture) in between. The unresolved seam is authentication — see harrison-sam-general-purpose-agents-2026 for the auth argument that plugs this stack into real enterprise work.