The agent boom got fragmented. Forge is the cockpit.
A working developer now runs three to five AI coding agents a day — a frontier model in the cloud, DeepSeek on a budget, a local model on the laptop, a specialist agent for reviews. Every one of them ships as its own island: its own terminal, its own session store, its own idea of status. The category's shared ceiling is transport — tools scrape terminal output and hook events instead of speaking a protocol, so status gets guessed from silence and sessions die with the window.
The gaps are specific and unclaimed. No tool uses ACP — the Agent Client Protocol — as its primary transport. No tool ships DeepSeek or local-endpoint support natively, even as local model servers became the default for cost-sensitive teams. The popular fleet managers are slow and put basic features — automations, org views — behind subscriptions. And the deepest problem is structural: the tools that delegate work to agents can't reach the code, and the tools that own the code have only a thin notion of delegation. Nobody ships both halves. That junction is exactly where Forge sits.
- No tool speaks ACP as primary transport — everyone scrapes terminals and hooks
- No native DeepSeek or local-endpoint harness anywhere in the category
- Automations and org features gated behind subscriptions as paywall bait
- Delegation tools can't touch the code; code tools can't delegate
- Heavy app shells for what is fundamentally a fleet of subprocesses
Sources: forge research corpus (2026-09) — deep dives read from the real source of six reference tools; feature matrix with adopt/reject decisions; internal dogfooding ledgers.