Coding with agents
My take IDE’s are mostly dead, cli agents might have peaked - server side not yet, and the shift for cli’s to servers, sandboxes, bot agents running on openai, cloudflare is visible. Developers lean to agent guis - desktop apps optimized for multi sessions, multi agent workflows. Hermes desktop and friends from major AI labs. Enterprises build their own platforms, meta harnesses, shared skills, mcps, connectors and observability. Everybody builds their own harness but that might not be the play given rate of improvement of models.
Vault kin (added at planting, not part of the verbatim take)
- Software engineering in AI era — the ghuntley JetBrains post-mortem (“they essentially Kodaked themselves”) is this note’s IDE-claim with the receipts; the Orosz talk’s IDEs-fading datapoint
- Meta harness — the enterprise lane named here: platforms, shared skills, connectors, observability (Omnigent = developer pole, Paperclip = executive pole)
- Pi agent harness — “everybody builds their own harness” — the eleven-harness landscape; championswimmer’s subagent-manager as the in-harness composition layer
- Agent sandbox — the servers/sandboxes shift (cloud sandboxes as first-class hosts)
- The open question carried in the last sentence — whether harness-building is a durable play “given rate of improvement of models” — is the bet openhuman makes the other way (in-process kernel, no harness accretion)- paseo (OSS) — the “agent GUIs, desktop apps optimized for multi sessions/multi-agent” prediction, shipped (20.1k★): daemon + full mobile client, worktree-per-agent, agents launching agents