Why We Structure AI Coding Around Three Modes — Not One Tool — And Cut Context-Switching Tax 40%

TL;DR: AI coding isn’t one workflow — it’s three distinct modes. We map every task to a mode first, then pick the tool. Terminal-autonomous for greenfield/infra, persistent chat-agent for long refactors/debugging, IDE-integrated for precision edits. No single tool wins all three. Full framework →

The Context

Two-dev team, 5 active codebases (monorepo + 4 client sites). Jan–Mar 2026 we treated “AI coding” as one activity, swapped tools weekly, lost 15–20% capacity to context switching. April 2026 we codified the Three-Mode Framework after tracking 200+ task outcomes. June 2026: Added fourth mode (inline completions) for Copilot.

What We Tested

Mode	Best Tool (2026-06)	Task Archetype	Anti-Pattern (What We Stopped Doing)
Terminal-Autonomous	Claude Code (cc)	Greenfield features, auth, infra, payments, test generation, migrations	Using for surgical TS edits (no LSP) or 5-min fixes (session overhead)
Persistent Chat-Agent	Codex (cx, ChatGPT Plus)	Multi-hour refactors, debugging, migrations, archaeological code reading	Using for greenfield (no terminal autonomy) or quick edits (chat UI friction)
IDE-Integrated	Cursor Pro (Composer OFF)	<30 min TypeScript surgical edits, type-error fixing, component extraction	Using for long sessions (Composer context loss) or autonomous loops (no allow-lists)
Inline Completion	GitHub Copilot	Boilerplate, imports, simple props, passive speed	Using for any multi-step reasoning (credits burn)

The Pivot Point

April 15, 2026: Stripe webhook migration (14 files, 3 hrs estimated). Assigned to Cursor Composer “because it’s in the IDE.” Result: 5.5 hrs, 3 context losses, 2 type regressions missed. Same task next week in Codex (Persistent Chat-Agent mode): 2.8 hrs, zero context loss, caught 2 type issues via manual review. The mode-task mismatch cost 2.7 hrs. We wrote the decision matrix that day.

June 3, 2026: Copilot AI Credits launched. Our agentic sessions burned $340 in credits in 2 weeks. Moved all multi-step work off Copilot. Added “Inline Completion” as fourth mode with strict guardrails.

What We Use Now

Decision Matrix (pinned in #dev-ai Slack, in .toolcrucible/decision-matrix.md):

If Task Is…	Use Mode	Tool	Trigger Phrase
New feature, 0→1, needs tests, infra	Terminal-Autonomous	cc	”greenfield” / “new endpoint” / “auth flow” / “infra”
Refactor >2 hrs, touches >5 files	Persistent Chat-Agent	cx	”migrate” / “refactor” / “debug” / “archaeology”
Type error, rename, extract component <30 min	IDE-Integrated	cursor	”fix type” / “extract” / “rename” / “quick”
Boilerplate, imports, simple props	Inline Completion	Copilot	(passive, always on)

Enforcement:

• PR template includes “AI Mode Used:” dropdown (Terminal-Autonomous / Persistent Chat-Agent / IDE-Integrated / Inline Only). CI fails if blank.
• Monthly retro reviews mode-task fit: gh pr list --json title,labels --jq '.[] | select(.labels[]?.name | contains("ai-mode"))'
• Team convention: cc = greenfield/infra, cx = deep refactors, cursor = precision, Copilot = passive.

Keybindings (.toolcrucible/keybindings.json):

• cmd+shift+c → cc in VS Code terminal (right pane)
• cmd+shift+x → cx in VS Code terminal (right pane)
• cmd+shift+v → Focus VS Code editor (Cursor LSP active)

When You’d Choose Differently

• Solo dev: Pick one mode, master it. Terminal-first → Claude Code only. VS Code loyalist → Cursor only (accept Composer limits). Chat-first → Codex only (accept no terminal autonomy).
• Team >5: May need 4th mode — “onboarding/standardization” → Windsurf (shared IDE config) or Cline (shared .clinerules prompts).
• Mobile/React Native: Expo integration makes IDE-Integrated mode higher value; might drop Persistent Chat-Agent.
• Strict local-only / air-gapped: Terminal-Autonomous → Aider + Ollama; Persistent Chat-Agent → not available locally yet; IDE-Integrated → Cursor/Windsurf work offline.
• Budget < $50/mo total: Cursor Pro $20 + Copilot $0 (personal) + BYOK API for terminal-autonomous ≈ cheapest stack.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	5	Framework eliminates tool FOMO; measurable velocity gain
Ease of Use	4	Requires team discipline; PR gate helps
Value	5	Zero cost (framework); tool costs unchanged
Support	N/A	Internal process; evolves with tool landscape

---

This is part of our AI Coding Tool Evaluation series. See full framework: AI Coding Workflow Modes: The Three-Mode Decision Matrix (2026 Edition)