Appendix A. For managers
10 min read
For the engineering manager, director or CTO who owns budget, staffing and procurement, and for the champion who has to hand them a case. Read Chapters 2, 8 and 9 for controls, task readiness and rollout. This appendix covers your decisions, an illustrative team, and a cost comparison you can replace with your own numbers. It can be forwarded on its own.
Your job in the 90 days
The manager owns the conditions for starting, the decision to expand, and the ongoing cost. Those responsibilities begin before the first pilot. The sample 90-day arc in Chapter 9 gives more attention to the manager later; it does not postpone procurement or governance until day 61.
Before starting, name the responsible owner and agree:
- The approved service and account, permitted repositories and data, and any required contractual terms.
- Initial permissions, secrets controls, isolation and recovery arrangements appropriate to the pilot. Verify the covered routes rather than assuming the configuration works.
- A spending limit, comparable work to observe, success criteria, and conditions for stopping or restricting the pilot.
- Who can judge the affected behavior, review consequential changes, and make the expansion decision.
During the pilot, protect time for investigation, review and maintenance. Help the lead resolve access, documentation and expertise gaps. A team unable to evaluate a payment algorithm needs qualified help or a narrower task; a subscription does not supply that expertise.
Later, decide whether the evidence supports expansion, a different workflow or stopping. Assign continuing ownership for instructions, integrations, vendor risk, controls and evaluation. These are operating responsibilities, even when the initial setup is complete.
Track outcomes from the start. Resist individual rankings based on PR counts, early promises of revenue, and a calendar that expands scope regardless of results. A small, well-understood pilot is easier to evaluate than a company-wide adoption target.
Which codebases now, later, never
Decide for the proposed task before sorting the portfolio. Define the affected behavior and environment. Establish a qualified evaluator, evidence that would expose a consequential mistake, and suitable permissions and recovery. If a prerequisite is missing or unknown, restrict that task to investigation or supervised preparation until the gap is addressed.
Chapter 8's eight signals then help prioritize preparation. The colors are a heuristic, not a delegation policy. With no unknowns, sum and round halves upward before applying a color:
| Rounded repository score | Preparation guide | Manager's decision |
|---|---|---|
| Green, 0-1 known gaps | Few preparation gaps | Consider bounded work whose task prerequisites are established. |
| Yellow, 2-3 known gaps | Substantial preparation | Fund the most consequential gaps and select work the team can evaluate now. |
| Red, 4+ known gaps | Broader preparation | Restrict the affected work while deciding what to investigate, improve, replace or retire. |
Unknown, ? |
Evidence is missing | Investigate it; do not silently count it as zero. |
A green repository score does not authorize every task. A high score does not prohibit useful read-only investigation. Missing qualified evaluation is a task prerequisite failure regardless of the total. Test coverage can inform the discussion, but relevant assertions, failure cases, reliable execution and the team's ability to interpret results matter more than a percentage.
Ask the lead for a short portfolio record: proposed task, prerequisites, known and unknown gaps, permitted scope, required preparation and owner. For each investment, ask what decision or verification capability it buys. "Never" should describe work the team cannot responsibly authorize under its present constraints, not a permanent label attached to a company or repository.
One team, week by week
Illustrative planning scenario: twenty engineers in a regulated financial-services firm, with an engineering manager and a cost-conscious CTO. This is not a measured pilot or a report of a client engagement. The schedule illustrates decisions and responsibilities; it supplies no observed adoption, cycle-time or defect-rate result. Use the sequence where it fits, and change the dates to match the evidence.
| Stage | What the team brings | What the manager does |
|---|---|---|
| Before the pilot | A bounded proposal, an accountable owner, approved account/data scope, initial controls, and an evaluation plan | Confirms eligibility, expertise, spending limit and stopping conditions before work begins. |
| First month | A small sample of accepted and failed attempts, review findings, billed usage and unresolved gaps | Protects evaluation time; addresses missing access or expertise; does not demand an adoption percentage. |
| Second month | A comparison of similar work, with human effort, delivery time, quality observations and costs recorded separately | Checks how tasks were selected and whether the results justify another bounded group of users or tasks. |
| Later manager phase | A recommendation to expand, change or stop, with limitations and operating responsibilities | Resolves licensing for the approved scope, assigns ongoing governance and maintenance, and reports supported outcomes. |
A useful board update names what was compared, what was observed and what remains uncertain. For example: "We compared [counts and task types] during [dates], using [selection method]. Human touch time was [result], elapsed delivery time was [result], and [defect observations] cover [window]. The incremental cost was [amount]. We recommend [decision], subject to [remaining uncertainty]." Fill the fields from records; leave unsupported claims out.
Keep one measurement card with the report:
| Field | What to record |
|---|---|
| Provenance and period | Observed pilot or illustrative example; dates, service/model versions, workflow and owner. |
| Work and comparison | Raw counts of attempted, completed and accepted tasks; task types, selection method, baseline/comparison group, exclusions and material differences. |
| Delivery time | Elapsed start-to-acceptance time, with the start/end definition and waiting time stated. Record human touch time separately. |
| Human effort | Investigation, implementation, review and rework for both groups; include failed or abandoned attempts. State how time was captured. |
| Quality | Escaped defects by severity, denominator and observation window. Record acceptance criteria and relevant checks; disclose missing follow-up. |
| Review | Reviewer effort, useful findings, resulting rework and required domain expertise. |
| Cost | Seats, all billed usage including failures, infrastructure, setup, ongoing maintenance and recurring governance/evaluation. State allocation and amortization assumptions. |
| Independence | Whether another engineer can operate the workflow and interpret its evidence without the champion. |
| Decision and limits | What supports expanding, changing or stopping; uncertainty, confounders and the next observation needed. |
Agent-touched PR share describes adoption. It is not an outcome target. A change in cycle time does not establish effort released, cash saved or unchanged quality. With a small sample or a short defect window, say what you observed; do not replace missing evidence with "within noise."
How much does agentic coding cost?¶
Count the full delivery cost, then compare useful accepted work. A seat quote is easy to find; investigation, review, failed attempts and maintenance are easier to miss. Prices and contract terms change. Use the structure below with your own agreement and records.
- Fixed seats. A genuinely fixed monthly fee stays the same when usage rises. Effective cost per useful task falls as useful usage increases. That does not make more generated output valuable. Usage caps, overages and tier changes are separate conditions to check.
- Metered usage. Count what is billed across all attempts, including retries, discarded work and additional agents. Use actual bills or usage records, not the final response's token count.
- Enterprise terms. Commitments, allowances, access controls, retention, audit facilities and contractual protections vary. Verify what the agreement provides and what the pilot requires; the label does not establish the terms.
Separate initial setup from continuing work:
| Cost | Include |
|---|---|
| Task delivery | Human investigation, implementation, review and rework, including failed attempts, on both sides of the comparison. |
| Tooling and infrastructure | Incremental seats, metered usage, environments and connected services. |
| Initial setup | Initial instruction, integration and control work; state the period over which it is amortized. |
| Ongoing maintenance | Updating instructions, skills, integrations and controls as the system changes. |
| Recurring governance and evaluation | Vendor-risk review, access review, measurement, control checks and procurement work that continues after setup. |
Keep three outcomes separate: shorter elapsed delivery time, useful engineering capacity released, and actual cash savings. A loaded hourly rate values capacity for comparison; it does not mean payroll falls when a task takes fewer hours. Record how any released time is used.
The structural levers¶
Change a cost lever only while checking the outcomes it must preserve. A cheaper run that misses the requirement is a failed attempt, not a saving.
- Effort and model choice. Compare settings on representative tasks. Use more reasoning or a stronger model where it changes the result enough to justify its cost; use a cheaper setting where the evidence supports it.
- Stable prefixes and caching. Where the provider supports prefix caching, stable instructions before changing task content can reduce repeated input cost. Check cache eligibility, actual hits and billed rates.
- Tool loading. Load the capabilities the task needs. Deferred schemas can reduce initial context where the client supports them; confirm that required tools remain discoverable.
- Context management. The cited JetBrains study compared observation masking and summarization with unmanaged history across five configurations. Both substantially reduced cost in most settings; masking was cheaper than summarization in four of five. The Qwen3-Coder 480B configuration saved 52.7% and 50.4% respectively against unmanaged history. This is not a universal 50% saving from replacing summarization with masking.
- Workflow and routing. Compare the native workflow with optional skills, reviewers or advisor arrangements on matched tasks. Include their startup, review and coordination costs. The right arrangement depends on the task and model.
Appendix C records the sources and their scope. Recheck the relevant result before treating it as a budget assumption.
A worked example¶
Hypothetical monthly comparison, not a measured result or vendor quote. A twenty-engineer team accepts twenty comparable pilot changes. The calculation assumes comparable useful outcomes for the arithmetic; it does not establish equal quality or prove that agents caused the effort difference.
Across those changes, the baseline uses 200 human task hours and the agent-assisted workflow uses 160. Both include investigation, implementation, review, rework and failed or abandoned attempts. Value human time at an illustrative loaded $100 per hour. Costs shared unchanged by both workflows are omitted from both sides.
| Monthly cost for 20 accepted changes | Baseline | Agent-assisted |
|---|---|---|
| Human task time: 200h / 160h at $100/h | $20,000 | $16,000 |
| Incremental seats | $0 | $530 |
| Metered usage, including failed attempts | $0 | $270 |
| Incremental infrastructure | $0 | $100 |
| Initial setup: 120h × $100 = $12,000, amortized over 12 months | $0 | $1,000 |
| Ongoing instruction/integration maintenance: 10h × $100 | $0 | $1,000 |
| Recurring governance/evaluation: 4h × $100 | $0 | $400 |
| Modeled monthly total | $20,000 | $19,300 |
| Modeled cost per comparable accepted change | $1,000 | $965 |
The maintenance and governance hours are additional operating work, separate from the 160 task hours. Task review and rework appear only in task time. Initial setup appears only in its amortized line. All attempts contribute their effort and usage costs, even though only accepted changes form the denominator.
The value calculation is:
$4,000 released task-capacity value − $900 tooling/infrastructure − $1,000 amortized setup − $1,000 ongoing maintenance − $400 recurring governance = $700 modeled net capacity value per month.
That $700 is not cash savings or a forecast. The initial setup consumes 120 hours upfront; amortization spreads its value for this comparison rather than changing when the work occurs. A different useful life, more review, fewer accepted tasks, worse quality or unused released capacity can change the result. Elapsed delivery time is not specified by these human-hour totals.
For an actual pilot, replace each assumption with a record or label it as an estimate. Compare the measurement card as well as the arithmetic. Cost per accepted task is useful only when the work and its accepted outcomes are comparable; twenty easy fixes and twenty consequential changes do not become equivalent because both sets contain twenty items.
Related: Appendix B.4 (task prerequisites and preparation worksheet), Appendix B.5 (example rollout calendar), Appendix B.10 (matched workflow comparison), Appendix C (sources).