Every AI coding agent has the same bad habit: dump the whole repo into the prompt and hope.
The result? Slow runs, burned tokens, and confident edits to the wrong files.
I got tired of it, so I built AI Workflow V3: a single Go binary that decides how much context an agent actually needs, before it writes a line of code.
The one rule: escalate capability, not context volume.
🧠 How it works
Give it a task, and it runs a deterministic pipeline:
Task
→ lane + risk classification (Answer / Small / Full)
→ hard context budget
→ repo-aware retrieval (index, BM25, structural, semantic)
→ evidence check (sufficient / requires_exploration / abstain)
→ bounded context → agent
→ verification → compact handoff
→ verified memory only
Your agent gets a tight brief instead of a haystack:
ai-workflow setup
ai-workflow brief "refactor payment retry" --format prompt
# agent works...
ai-workflow verify --check "go test ./..." --strict
ai-workflow memory add --type verified-fix --keywords "payment retry" \
--summary "retry uses jittered backoff" --file svc/retry.go
⚡ What makes it different
| 📦 One native binary | No Python env. Windows, Linux, macOS, WSL. |
| 🔌 Built-in MCP server | Plug into Claude Code, Cursor, Codex, Windsurf. |
| 🗂️ Multi-repo discovery | Finds nested Git roots and routes retrieval across them. |
| 🚦 "Should I even edit yet?" | Evidence states stop agents from guessing when context is thin. |
| 🔒 Hardened by design | No shell execution, bounded output, timeouts, constrained env. |
| 🧾 Replayable runs | Hash-chained run journals show why every decision was made. |
| ✅ Memory you can trust | Only verified fixes are saved. No junk in, no junk out. |
| 🛡️ Supply-chain safe | SHA-256 verified installs + GitHub build-provenance attestations. |
🔌 Connect your agent in one line
claude mcp add ai-workflow -- ai-workflow --root /path/to/project mcp
Your agent then calls brief, context, route, search, memory_search, verify, and more natively, with no shelling out.
🔁 Why a V3?
V2 was Python and proved the idea. V3 is a clean Go rewrite, and it doesn't get to call itself compatible just because features were reimplemented. Every V2 behavior counts as migrated only after a differential fixture proves the same input gives an equivalent result.
🚀 Try it in 10 seconds
curl -fsSL https://raw.githubusercontent.com/Taki7980/Ai-workflow-V3/main/install.sh | sh
Windows:
irm https://raw.githubusercontent.com/Taki7980/Ai-workflow-V3/main/install.ps1 | iex
⭐ Repo: https://github.com/Taki7980/Ai-workflow-V3
MIT licensed. Stars, issues, and brutal feedback welcome.
How do you keep your agents' context under control? Drop your approach below 👇
