Appendix A. For managers
9 min read
For the engineering manager, director or CTO who owns budget, hiring and procurement - and for the champion who has to hand them a case. Read Chapters 2, 8 and 9 first (guardrails, readiness, rollout). This appendix holds the rest: your role, the portfolio call, one worked team, and the money. It is built to be forwarded on its own.
Your job in the 90 days
The manager does not run the agent. The manager defends three boundaries.
The manager phase is days 61-90, after the champion has stood up the practice and the lead has hardened it (Chapter 9 has the arc). It has three centers of gravity:
- Procurement. Close the licensing decision so the team is not on individual seats forever.
- Governance. Sign off on the security posture the champion drafts.
- Reporting. Translate operational results into terms leadership will fund.
None of it needs you to write code. All of it needs you to hold lines: operational data versus ROI promises, tooling matched to usage versus forced uniformity, protecting the team from premature metrics versus demanding evidence that the work pays off.
You also own the calls nobody else can make: hiring the cryptographer behind a signal-six problem, funding a quarter to move a yellow project to green, buying extra tooling or services.
What you do not do:
- Chase per-engineer productivity metrics in month two.
- Commit to a three-year ROI before the dashboard has a quarter of data.
- Authorize vendor-versus-vendor benchmarks before the team has shipped enough to define what it is benchmarking.
Those questions destroy adoption when answered too early. In months one to three, shield the team from them, and tell stakeholders up front that the first thirty days show almost no measurable productivity. When the operational number is solid, defend it; leave the revenue argument to whoever owns revenue. By day 90 your role shifts from standing up the practice to tracking the metrics each quarter.
Without an engaged manager the cost shows up elsewhere: procurement sits two months waiting on legal, the documentation budget is never allocated, yellow projects stay yellow. The lead's classification was correct. It had no buyer.
Which codebases now, later, never
Apply the traffic light per project, never per company. The question is not "is this company ready for AI." It is "which of our codebases is ready today, and what would it take to move the others." Chapter 8 holds the eight kill signals and the scoring. The part you act on is the sort.
| Color | Signals | Mode | Your call |
|---|---|---|---|
| Green | 0-1 | Agent-led | Now. Start here to build the team's experience and confidence. |
| Yellow | 2-3 | Human-led, agent support | Later. Fund the structural fixes that move it toward green over a quarter or two. |
| Red | 4+ | Fix the codebase first | Never, as it stands. Reform or retire over a longer horizon; the agent stays out until the signals close. |
Most companies hold a mix, and the mix is the plan. When leadership has decided to "adopt AI," the temptation is to put the agent on a red codebase anyway. Resist: that temptation is what produces the failed adoption stories that taint the field.
Two rules to know before you read a scorecard:
- Signal six is weighted more heavily than the others. It is counted with them and also overrides the count: a codebase that scores zero on signals one through five but yes on signal six is still red for the relevant work. The fix is structural: build the capability, partner with someone who has it, or restrict the agent away from those areas.
- Count by intensity, not presence. "Tests exist but are partial" is half a signal. A codebase that technically has tests that cover 10% of the code at low quality is not green.
Ask your lead for a one-row-per-project scorecard on the top five active projects: signals scored, color, recommended mode (agent-led, human-led, fix first). It takes about two hours. For a yellow project, ask for the two or three structural fixes that would make it green. For a red one, ask whether the answer is investment, replacement or retirement. Sometimes the question is the strategy, not the codebase: a vendor-customized fork with severe upstream coupling asks whether the custom layer should exist at all.
One team, week by week
Twenty engineers in a regulated financial-services firm. The manager owns the engineering function and reports to a supportive but cost-conscious CTO. The board wants a one-page update at the end of the third quarter on the AI investment. The champion ran days 1-30 and the lead hardened the practice in days 31-60; the manager picks up from day 61.
| Week | What lands | What the manager does |
|---|---|---|
| 1 (days 61-67) | A dashboard of three metrics: share of merged PRs touching agent code; cycle time and defect rate on that set, each compared to the prior quarter. | Verifies the metrics are real. Does not try to grow them. |
| 2 (days 68-74) | Procurement asks seat versus enterprise. The team is on individual Pro seats. | Chooses Team tier for 13 of 20 engineers and Pro seats for the 7 who rarely use the agent. Defends it: the goal is bounded spend that matches usage, not uniform tooling. |
| 3 (days 75-81) | The security committee has read about the configuration-file injection class Check Point Research disclosed in early 2026 (Appendix C) and wants a written governance posture. | Edits and signs a one-page document naming four control points: hooks, sandbox, secrets vault, telemetry (Chapter 2 has the layers). The champion writes it. |
| 4 (days 82-88) | The board ask. | Sends two sentences and one number. Promises no revenue figure. |
The numbers behind week 1: 41% of merged PRs agent-touched in month two, cycle time on them 28% below the pre-agent baseline, defect rate within noise of it.
The board update, in full: "We have twenty engineers on agent-assisted delivery. Cycle time on agent-touched PRs is 28% lower than baseline at no measurable change in defect rate. We are continuing to scope expansion through the next quarter." The manager defends the operational number and leaves the revenue argument to the CTO.
Day 90: back to ongoing operations. The dashboard stays; the champion and lead keep their roles.
How much does agentic coding cost?¶
The sticker is the smallest line, and the one everybody asks about first. Prices change every quarter; the rubric does not. Plug your own numbers into it.
Three pricing models dominate:
- Per-seat. A flat monthly fee per engineer. Predictable; cheap for occasional users, expensive for power users who run the agent all day.
- Per-token. You pay for what the agent reads and writes. Fair to occasional users; punishing for the engineer running three parallel sessions for ten hours.
- Enterprise. A negotiated annual commitment bundling seats, a token allowance and the compliance surface (Zero Data Retention, audit log access, dedicated capacity, SSO, signed BAAs). It amortizes both extremes and adds what regulated industries require.
The bounding heuristic. Per-engineer cost is bounded by what the team would otherwise spend on similar tooling (IDE licenses, code-intelligence platforms, AI-coding subscriptions) plus a fraction of a senior engineer's hourly time saved per week. If the all-in monthly cost exceeds that envelope by a wide margin, the math is unlikely to work with any vendor. If it falls comfortably inside, it is unlikely to fail. Break-even lands quickly when the agent saves even a few hours a month per engineer.
What the quote leaves out. Four lines dominate the real total:
| Line | What it covers | The 20-engineer team (illustrative) |
|---|---|---|
| Seats (the sticker) | The vendor quote | ~$530/mo, ~$6,400/yr, recurring |
| Integration | Custom skills, hooks, MCP servers for internal systems | A few engineer-weeks, one-time |
| Skill-authoring | Maintaining AGENTS.md, updating skills as the codebase evolves | A few hours per engineer per month, plus the champion's part-time quarter |
| Review | Reviewing agent output: less per change than hand-written code in most cases, never zero | Concentrated on senior reviewers, ongoing |
| Governance | CISO security review, Zero Data Retention addendum, procurement cycle, audit logging, vendor-risk monitoring | A week to a quarter, front-loaded in a regulated firm, one-time |
The structural levers¶
Prices go stale; the levers that move the token line do not. On per-token plans they are where the second-year savings live.
- Effort level. The Chapter 1 primitive and the largest cost dial you own. Set it per phase, sweep it per model, and never leave the default on for the whole loop.
- Cache-first prompt layout. Stable content first (system prompt, AGENTS.md, loaded skills), per-turn content last, so the prefix caches. Across 500 sessions on three providers that ordering alone cut cost by 41% to 80%.
- Deferred tool loading. One audit of a default Claude Code session counted about 24k tokens of tool definitions and schemas before the user's first word. Defer the schemas and pay only when a tool is used.
- Masking over summarization. Blank old tool results instead of summarizing the history. JetBrains measured it over 50% cheaper, with a solve rate that matched or beat summarization.
- Tier routing. Route by task, not by team: a cheaper model does the work, a stronger one stays on call. In Anthropic's advisor pattern (April 2026), Sonnet with an Opus advisor beat Sonnet alone at 11.9% lower cost per task; Haiku with the advisor was 85% cheaper than Sonnet alone but trailed it by 29% in score. Right for the mechanical tiers of the loop, wrong for research.
A worked example¶
Run the 20-engineer team through the rubric. The figures are round numbers for the arithmetic, not quotes.
Sticker first. 13 Team seats at an illustrative $30 plus 7 Pro seats at $20 is $530 a month, call it $6,400 a year. Write it down: it is the smallest number on the page.
Then the human lines. A few engineer-weeks of integration, at any realistic loaded cost, clears the first year's sticker on its own, once. Skill-authoring across 20 engineers, plus the champion's quarter, recurs and can run larger than the seats themselves. Review lands on your most expensive hours, the senior reviewers. Governance in a regulated shop is a week to a quarter of security, procurement and audit-logging work before a single seat is billed.
The value side runs the bounding heuristic in reverse. At an illustrative $100 loaded per hour, a $30 seat is covered when the agent saves that engineer about twenty minutes in the month; a few saved hours is nowhere near close. But that back-of-envelope is not what justified the spend. The operational number did: 41% of merged PRs agent-touched, cycle time 28% below baseline, defects flat. Measured hours out, not projected dollars.
The discipline: the four hidden lines dominate, and the human ones dwarf the seat math. Defend the operational number to the board, not a projected ROI the dashboard cannot yet support. Walk into procurement with your own numbers in this structure. Pricing changes; this arithmetic does not.
Related: Appendix B.4 (kill-signal scoring worksheet), Appendix B.5 (90-day calendar), Appendix C (sources).