Why We Built Our Own AI Agent Tooling Layer — And Why Every Team Will

TL;DR: Three-dev team built a 4-layer agent tooling layer (MCP servers → agent-native git → local knowledge base → persistent context). Agents now drive 70% of terminal, git, test, deploy operations. Off-the-shelf tools assume humans drive; the layer inverts this. Full comparison →

The Context

Three-dev B2B SaaS team, agent-assisted since Jan 2026. Baseline: human prompts → agent writes code → human commits/pushes/tests/deploys. Friction: human bottleneck on every non-coding operation. Agents couldn’t use internal tools (deploy CLI, schema validator, sync engine) — only REST/CLI built for humans. June 2026: built tooling layer agents can actually use. Now: human defines intent → agents execute full loop (code + git + test + deploy + monitor). Human reviews intent.md + risk.md only.

What We Tested

Layer	Off-the-Shelf Approach	Our Agent Tooling Layer	Verdict	Why
Tool Access	Agents call REST/CLI (hallucinate params, parse output)	MCP servers (stdio) for every internal tool	✅	Typed interfaces, structured errors, 95% agent success vs 35%
Git Operations	Human drives branch/PR/merge; GitHub Actions CI	Cursor Origin (agent-native git hosting)	✅	Agents own branch→PR→review→merge; human reviews intent only
Knowledge/Memory	Agents re-read docs/files every session (context window)	Local ChromaDB + MCP (persistent RAG)	✅	Agents query ADRs, API specs, error patterns, decisions across sessions
Persistent Context	Context lost on session reset, model switch, memory pressure	Continue.dev (Zed) + Ollama (terminal/DB/file state persists)	✅	4-6 hour sessions with live tunnel, DB connection, full file context
Observability	Human checks Datadog; agent can’t query	pino + Loki + MCP (agent queries logs/metrics)	✅	Agent debugs own failures: “Check logs for 5xx last 10 min” → structured response

The Pivot Point

April 2026: Asked agent to “deploy preview for auth-refactor branch.” Tried 4× via REST — wrong endpoint, malformed payload, token expiry. Built minimal MCP server for deploy tool (150 lines). Same prompt next day: agent called mcp__deploy__create_preview({branch: "auth-refactor"}), got structured response with URL + log tail. Deploy succeeded in one shot. That week we MCP-wrapped 3 internal tools. Agent tool usage went from 0% to 60% of team invocations in 14 days. The pattern: every operational tool the team uses daily needs an MCP endpoint. Not “someday” — now.

What We Use Now (4-Layer Agent Tooling Stack)

Layer 1: Tool Access          → MCP servers (stdio) for deploy, schema, sync, logs
Layer 2: Git/Collaboration    → Cursor Origin (agent-native branches, PRs, reviews, merges)
Layer 3: Knowledge/Memory     → ChromaDB + MCP (ADRs, API specs, error patterns, decisions)
Layer 4: Persistent Context   → Continue.dev (Zed) + Ollama (terminal state, DB connections, files)

Agent workflow (typical feature):

1. Human creates intent.md in spikes/<feature>/ — what, why, constraints
2. Planner agent (qwen2.5-coder:32b) reads intent + ChromaDB context → creates plan
3. Executor agents (deepseek-coder:33b) implement: code → test → mcp__schema__validate → mcp__deploy__create_preview
4. QA agent runs integration tests in Origin sandbox → posts results to agent PR
5. Security agent reviews diff → comments on mcp__schema__diff risks
6. Human reviews intent.md + risk.md summary → clicks “Approve Intent”
7. Origin merges → mcp__deploy__promote_to_staging → monitoring alerts via MCP

Human time: 3-5 minutes. Agent time: 20-45 minutes. Zero human git/terminal ops.

When You’d Choose Differently

• GitHub Copilot Workspace if: org mandates GitHub, not ready for Cursor Origin, compliance needs GitHub audit trail
• GitLab Duo if: self-hosted required, enterprise RBAC, existing GitLab investment
• Custom agent framework (LangGraph, AutoGen) if: complex multi-agent orchestration beyond coding, non-coding workflows
• No tooling layer if: team <3, low agent adoption, human-driven workflow acceptable, compliance blocks agent ops

Tool Crucible Rating (Layer Aggregate)

Dimension	Rating (1–5)	Notes
Overall	5	Highest leverage investment in 2026; compounds with model improvements
Agent Autonomy	5	70% of ops agent-driven; human only on intent/ambiguity
Build vs Buy	4	MCP servers ~150 LOC each; Origin waitlist; ChromaDB + MCP ~200 LOC
Maintenance	3	4 MCP servers + ChromaDB + Origin sync — manageable but growing
Portability	3	Origin = Cursor lock-in; MCP/ChromaDB/Continue portable across editors

---

This is part of our AI Agent Tooling series. See full comparison: AI Agent Tooling Layer 2026: MCP + Origin + ChromaDB + Persistent Context — Build Guide