Ten engineers, the output of thirty: the honest enterprise agentic playbook

A vendor will tell you that handing your developers an AI assistant turns ten of them into thirty. A skeptic will tell you it turns them into ten developers who write worse code faster. Both are wrong, and the gap between them is where the actual decision lives.

I have spent twenty-five years watching tools get sold as multipliers. Most were not. This one is, on a specific slice of the work, if you build the system around it. The trouble is that the system, not the model, is the hard part, and the system is the part nobody puts on the invoice.

Where should AI coding agents run? Local, self-hosted, frontier, or managed cloud

The first mistake is asking "which AI do we buy." The right question is "where does each task run." A serious organization runs four tiers at once and routes each request to the cheapest one that can do it well.

The economics are the reason this matters. A 32B-class open model such as Qwen, served on one well-utilized high-memory card, costs roughly twenty to forty cents per million tokens. A frontier model costs around fifteen dollars for the same million. On token price that is forty to eighty times cheaper; once you count the operations team, power, and the frontier vendors' prompt caching, the honest all-in gap is closer to eight to fifteen times. Still large, and your code never leaves your network. (Size the model to the hardware: a 600B-class model like DeepSeek needs a multi-GPU node, not a single card.) So the goal is not to run the cheap model everywhere, which caps your quality, nor the expensive one everywhere, which empties your budget. The goal is a router that knows the difference.

One honesty note on the title. The pitch says thirty; the number you can defend to a board is closer to two. A well-run team of ten reads like fifteen to twenty across a normal portfolio, and only approaches twenty-five to thirty on narrow, mechanical work. Thirty is the ceiling on the easy slice, not the average. The rest of this piece is about earning the honest 1.7 to 2.1 times, because that is the number that survives the first skeptical CTO.

Which task goes to which tier is the routing decision in one table. This is the matrix the Agency Router applies on every request.

What governance and security do AI coding agents need?

This is the question that decides whether the rollout survives the security review, and it is the reason the self-hosted tier exists. Agents that read your repositories and call tools are a new attack surface and a new path for intellectual property to leak. Before a single agent touches production, the CISO needs four answers, enforced in the architecture rather than in a policy memo.

• Data residency and IP. Pin sensitive repositories to the local or self-hosted tier; redact secrets at the gateway; verify no-train terms in every vendor contract; never allow proprietary code into a consumer chatbot. The privacy order, most protective first: local, self-hosted internal, private cloud endpoint, API with a no-train agreement.
• Quality as a measured number. An evaluation harness built from 50 to 100 real tickets in your history, run on every model or router change. Without it you are governing on vibes, and vibes do not survive an incident.
• Cost containment. The router enforces budget caps; cheap is the default; the runaway ceiling is alarmed; the routing split is a live dashboard.
• Accountability. If an agent ships a vulnerability, a named human owns the routing decision and the merge, not just the code. A human gate stays on production.

When agents lose money, and what happens to your engineers

An honest guide names the cases where this fails. Agents lose money on familiar code worked by senior engineers, where the review-and-correct loop costs more than the help; on tangled legacy with no fast tests, where the agent gets lost and confidently ships defects; and any time leadership budgets for a flat 3x and meets a 1.6x reality on maintenance work. The fix is always the same: fix the prerequisites before you scale, and route those tasks to a human.

The workforce question is the one a CEO asks next. This does not empty the org chart; it changes its shape. Your seniors stop spending afternoons on boilerplate and spend them on architecture, review, and judgment, which is where the agents cannot follow. Juniors need deliberate protection: rotate them through agent-off work and make review a teaching moment, or you raise a generation that never learns to debug. The new roles are the platform engineers who own the router and the evaluation harness, and they are the hire that actually pays for itself.

Do AI coding agents really make teams faster? The honest multiplier

The gains are not uniform, and the flat "3x across everything" is marketing. Boilerplate, scaffolding, test generation, and mechanical refactoring compress dramatically. Novel architecture and tangled legacy barely move. On familiar code worked by senior engineers, a careless rollout can even land below 1.0, slower than no agent, when the review loop costs more than the help. Your blended figure depends on your portfolio; for a real mixed codebase it lands between 1.7 and 2.1 times after five or six months of disciplined work. The slice figures below are what we measured on real pilots, not a vendor's chart.

And the multiplier is gated by the codebase, not the model. An agent loses its way in a tangled repo the same way a new hire does, only faster and more confidently. Before you count a single hour saved, you pay for codebase hygiene, fast and trustworthy tests, disciplined review, real guardrails, and hardware that does not throttle your developers. None of that is on the vendor's invoice. All of it is on yours.

What a CEO should ask before signing

If the plan in front of you cannot answer those four, it is a vendor's plan, not yours.

The full field guide

This post is the on-ramp. The complete strategy, with the four tiers and the decision tree, the real cost models, the privacy ranking, the six-month roadmap, the seven failure modes, and a one-page cheat-sheet a CIO can hand straight to a CEO, is a free field guide. Open it in the reader, or take the PDF.