How We Structured Our AI Coding Agent Workflow — Three Tools, Three Modes, Zero Context Switching Tax

TL;DR: Forcing one tool to do everything creates friction. We route by task type: cc (Claude Code) for terminal autonomy — refactors, tests, infra, deploys; cx (Codex) for research, scaffolding, docs; cursor for <30 min TS/React surgical edits. Unified cost dashboard + weekly review keeps spend at $140/mo (down from $287). Full workflow spec →

The Context

Two-dev team, ~120 hrs/mo AI coding across Basso Digital OS + 5 client sites. Jan–May 2026: tried to make Cursor do everything — Composer for refactors, chat for questions, inline for completions. Result: context loss on long tasks, no terminal autonomy, unpredictable Anthropic API bills ($150–300/mo). June 2026: adopted task-type routing after Anthropic revealed 80%+ of their codebase via Claude Code terminal agent.

What We Tested

Workflow Pattern	Tools	Verdict	Why
Single-tool (Cursor Pro)	Cursor for all	❌	Composer context loss; no test/deploy autonomy; $200 Ultra gate
Single-tool (Claude Code only)	Claude Code for all	⚠️	Overkill for quick edits; no persistent research agent mode
Single-tool (Codex only)	Codex for all	❌	No terminal autonomy; different UX for editing
Task-type routing (current)	cc / cx / cursor	✅	Right tool per mode; $140/mo predictable; zero context tax

The Pivot Point

May 28, 2026: Debugging a Stripe webhook idempotency bug. Cursor Composer: 4 context resets, couldn’t run stripe listen to test. Switched to cc (Claude Code): wrote test, ran stripe listen --forward, fixed race condition, ran migration, deployed to staging — single session, zero resets. Same day: needed to research Stripe Connect onboarding patterns. cx (Codex): persistent 2-hr research session, generated scaffold, saved to Notion. Realized: different tasks need different agent UX paradigms.

What We Use Now

Routing Protocol (enforced via aliases + team discipline):

# ~/.zshrc — source of truth for routing
alias cc='claude-code --model sonnet'                           # Terminal autonomy mode
alias cc-opus='claude-code --model opus --reason'               # Opus: auth|infra|payments only
alias cx='codex'                                                # Persistent research/scaffold
alias cursor='cursor'                                           # Surgical TS/React <30 min

Task Type	Trigger	Tool	Example
Architecture / Multi-file refactor	cc	Claude Code	”Refactor auth module across 12 files, add tests, run migration”
Infra / Deploy / CI	cc	Claude Code	”Update GitHub Actions, deploy to staging, verify health checks”
Debugging with reproduction	cc	Claude Code	”Reproduce Stripe webhook race, write test, fix, verify”
Research / Learning	cx	Codex	”Research React Server Components patterns, scaffold demo”
Scaffolding / Boilerplate	cx	Codex	”Generate Next.js + Supabase + Tailwind starter with auth”
Documentation / ADRs	cx	Codex	”Write ADR for payment provider switch, update README”
Quick TS/React edit	cursor	Cursor Pro	”Fix type error in Button.tsx, update Storybook”
Inline completion	VS Code	Native TS LSP	Daily typing — no AI marginal cost

Weekly Review (Monday 9am, #dev-ops Slack):

1. claude-code usage --since 7d --by-model → Opus hours × $15 = Opus spend
2. If Opus > $30/wk: audit --reason flags; move tasks to Sonnet + better prompts
3. Cost dashboard projection: ~/.toolcrucible/cost-dashboard.sh
4. Routing friction check: any tasks misrouted? Update aliases/protocol

Cost Dashboard (runs daily 6pm via cron):

#!/bin/bash
# ~/.toolcrucible/cost-dashboard.sh
CC_USAGE=$(claude-code usage --json --since "$(date +%Y-%m)-01")
CC_PROJECTED=$(echo "$CC_USAGE" | jq '.projected_monthly_usd')
TOTAL=$((CC_PROJECTED + 20 + 20))  # Cursor + ChatGPT Plus fixed
echo "[$(date)] CC_Proj: \$${CC_PROJECTED} | Total_Proj: \$${TOTAL}" >> ~/.toolcrucible/cost-log.csv
if [ "$TOTAL" -gt 150 ]; then slack-post "#dev-alerts" "💰 AI coding projected: \$${TOTAL}/mo"; fi

When You’d Choose Differently

• Solo dev, <20 hrs/mo: Two-tool max (Cursor + Codex, or Claude Code only). Three-tool routing overhead not worth it.
• Team >5: Per-seat credit pool ($100×) may exceed enterprise API discounts. Standardize on 1–2 tools.
• No terminal comfort: Cursor/Windsurf dedicated editors lower friction; accept context loss tradeoff.
• Budget = $0: Aider + Ollama (local) + VS Code Continue (free) + Copilot Free tier. No persistent agent equivalent.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	5	Routing protocol = workflow clarity; 51% cost reduction real
Ease of Use	3	Alias discipline required; onboarding doc essential
Value	5	$147/mo saved = 1.5 client projects/yr
Support	N/A	Internal protocol; evolves with pricing changes

---

This is part of our AI Coding Tool Evaluation series. See full comparison: AI Coding Agent Routing Protocol 2026