P3 Run & Learn: how to run an AI agent in production safely

In this lesson you'll learn:

• how to cut over from shadow to full traffic without betting on a date;
• how to catch drift, the defect with no error message, before a customer does;
• how to report value and cost so the programme survives its first bill.

Sound familiar?#

• Launch day was set in a meeting, and nobody divided the cases you need by the cases you get a day.
• Three months after launch a customer noticed the assistant now offers credits where it used to offer refunds.
• Finance found the token bill before the product manager reported the saving.

P3 is where each of those is either caught by a mechanism or discovered by someone else.

What is P3 Run & Learn?#

P3 is the fourth phase of the agentic PDLC, and the one that never finishes. The sponsor is accountable for it. DevOps runs the flags, traces and recovery; QA watches drift and runs the injection suite; the product manager reports the two numbers; the architect turns incidents and bills into design changes.

A P3 cycle ends when both numbers are reported and the brief for the next P0 exists.

P3, step by step#

Step 1 · Widen on evidence, not on a date#

Four states, taken one at a time: shadow, where the agent decides on real traffic and acts on nothing; a 5% canary; wider, on evidence; then all of it. Each step is a condition, never a date, and the arithmetic decides how long it takes. At 240 cases a day, 5% is twelve a day, so five hundred cases takes 42 days. The safe share is the slow one, which is exactly why a cut-over widens rather than holding.

Money actions stay gated at every stage, whatever the shadow shows. When the SkyWays sponsor had six weeks instead of ten, the shadow window was cut from fourteen days to seven, on the simple slice only, with every disagreement read.

Step 2 · Rehearse the way back, and cap the runaway#

Rollback is a capability, and a capability nobody has used is a belief. Before cut-over, throw every switch with a stopwatch running and write the times down. SkyWays measured four: the kill switch in 40 seconds, flag to shadow in 2 minutes, a prompt rollback in 3, and a model rollback in 11, because it redeploys the runtime. Then cap the runaway: a loop limit on every agent, and a cost limit per case that ends in a person's queue rather than in another retry.

Step 3 · Watch for drift#

Drift is behaviour changing with no deploy, no error and no alert. Your existing monitoring answers whether the system is up and fast; it was never looking for this. Watch the output mix (the share of each kind of decision), because accuracy needs labels and arrives late, while the mix is visible the same day.

Watch it against two thresholds, not one. The SkyWays refund-to-credit mix slid from 61/39 in week one to 48/52 in week eight, thirteen points, and never tripped its 5% weekly alert, because it moved 1.9 points a week. A second threshold, on the level against a frozen baseline, fires within weeks where the weekly one never fires at all. Wire a breach to re-open the release gate automatically.

Step 4 · Report two numbers on one line#

Report what the feature saved and what it cost, together, every cycle, before anyone asks, with the two rows that stop either number being gamed: the review hours added, and the re-runs.

A first cycle that saves time and costs more is survivable, if the sponsor hears it from you. On day ninety SkyWays reported 40 to 45 percent fewer person-days and a token bill of $4,200 on one line, with the rising review hours beside them and the reason they would fall. The programme continued, not because the numbers were flattering, but because both of them came from the team.

Step 5 · Turn what you learned into the next P0#

A bill that leaves its estimate is a design question, not a finance one. SkyWays' day-75 bill was 4.4 times its estimate on flat traffic, and a per-call log showed four ordinary habits multiplying:

An incident is the same: its postmortem asks which enforced control was missing, never who was careless, and it ends as a typed parameter, a test and new golden cases, plus a brief for the next P0. The eight loops are how that learning finds its way back.

Where you'll use it#

• From the first day of live traffic, forever. P3 does not end; each cycle ends.
• After every prompt, model or configuration change, which re-runs the harness and, if the mix moves, re-opens the release gate.
• In every sponsor update, which carries the two numbers on one line.

Why it matters#

Gartner's forecast that over 40% of agentic AI projects will be cancelled by 2027 names escalating costs as a cause. Costs escalate quietly and are found by finance; drift degrades quietly and is found by customers. P3 exists so that both are found by you, first, with a design change already drafted.

Key takeaways#

1. Widen on evidence: shadow, 5%, wider, all, each step a condition, and its length set by arithmetic.
2. Watch the output mix against two thresholds, and let a breach re-open the release gate on its own.
3. Report two numbers on one line, the saving beside the spend, and turn every bill and incident into the next P0.

FAQ#

How do you monitor an AI agent in production?#

Add three signals to your usual uptime and latency monitoring: the cost of each case, a trace of what the agent did and why, and the mix of its outputs over time. Alert on the mix against both a weekly change and a fixed baseline, and review a sample of cases against human judgement on a schedule.

What is AI model drift?#

Drift is a change in an AI system's behaviour without any change to its code, because the world, the inputs, the upstream model or the data it retrieves has moved. It raises no error, which is why it is watched through the distribution of outputs rather than through failures.

How do you roll back an AI agent?#

Keep four switches and time each one in a rehearsal before launch: a kill switch, moving the action's flag back to shadow, rolling back the prompt, and rolling back the model version. A model rollback is usually the slowest, because it redeploys the runtime.

How do you measure the ROI of an AI agent after launch?#

Report two numbers together: the saving, measured against the baseline from P0, and the full cost, including tokens, review time and re-runs. Reporting only the saving is how a programme gets cancelled when finance finds the other number.

Apply it in your role#

Across the enterprise. Aggregate every product's two numbers and drift chart into one portfolio view. The incidents and bills P3 produces are the pipeline of next quarter's P0 briefs.

The ten-minute workflow. The cycle report, drafted from raw data:

Here is this cycle's data: <tracker export, token log, review log>. Write the two-number report:
person-days per story against the baseline of <n>, token spend per story, review hours added and
re-runs per story, each with its change. Add one paragraph on the trend and the single biggest risk
to the saving. Do not smooth away a rising re-run count.

Sources and credits#