IPRE Pipeline

Definition

The IPRE Pipeline is the four-stage process that closes the gap between a strategy a human can state and a strategy a Machine Core can act on: Intent → Plan → Run → Evaluate. Intent is the human expression of the outcome wanted, precise about priorities and constraints. Plan is the translation of that intent into machine-executable structure, reviewed and approved by the Human Cortex before anything runs. Run is autonomous execution by the agent layer within the approved parameters. Evaluate closes the loop — evaluator agents monitor outcomes against the original intent, catch drift toward proxy optimization, and surface what the Cortex needs to revise Intent in the next cycle. It is strategy-as-code operationalized: the recurring heartbeat that keeps a fleet of agents legible to the few humans directing it.

Why it exists / the problem it solves

Most organizations have a strategy. Very few have a strategy their systems can act on. The problem is translation. A CEO says: "grow enterprise revenue 40% this year while maintaining margin above 18%." That sentence is clear to humans and inert to agents. Agents need a precise definition of "enterprise revenue" — which segments, which contract types, which geographies — plus measurable leading indicators they can optimize in real time, explicit constraints on what must never be sacrificed in pursuit of the goal, and a success definition sharp enough that the system can tell whether a candidate action advances the objective or violates it.

Left untranslated, that gap does not stay empty. Agents fill it with their best guess at what you probably meant — and Goodhart's Law governs the guess: they optimize what is measurable, not what matters. The IPRE Pipeline exists to perform the translation deliberately, stage by stage, with a human checkpoint placed exactly where it does the most good and removed everywhere it would only slow the system down.

Anatomy

Intent. The human expression of what the organization wants to accomplish — the outcome, not the method. Expressed in natural language, but precise about priorities, constraints, and acceptable trade-offs. "Grow enterprise revenue" becomes: "Increase ARR from enterprise customers (>500 employees) by 40% before December 31, within a gross-margin corridor of 18–24%, prioritizing three-year contracts over annual." Intent is where the Human Cortex declares what success means.

Plan. The translation of intent into machine-executable structure. A planning agent decomposes the intent into specific objectives, identifies the [[vp-agent-architecture|VP-Agents]] responsible for each, defines the KPIs each will optimize, and specifies the constraints — the lines the system must not cross even in pursuit of the goal. This is where the Human Cortex reviews and approves before execution begins. It is the last purely human checkpoint before the system runs autonomously.

Run. Execution. VP-Agents dispatch specialists, coordinate across functions, allocate resources, and respond to real-time signals — all within the parameters set in Plan. The defining discipline: the system does not improvise the objective. It improvises the method.

Evaluate. The loop's close. Evaluator agents monitor outcomes against the Intent definition, detect drift toward proxy optimization (winning the metric rather than the goal), flag anomalies exceeding a VP-Agent's charter authority, and surface the signal the Cortex needs to adjust Intent next cycle. If the Evaluate stage produces only green dashboards, something is wrong with the evaluation criteria, not the operation.

The single most important seam is the Plan-to-Run boundary — the last moment human judgment is applied before the system executes on its own. Get it right and you have minutes of review preventing months of post-mortem. Get it wrong and the damage accumulates at machine speed before evaluation catches it.

Figure: What happens after the Plan checkpoint. Once Intent is approved, agents execute autonomously through their own perceive-reason-act-learn cycles during Run, and their outcomes feed the Evaluate stage that closes the IPRE loop back to Intent.

How it works in practice

Chapter 5's cautionary tale lives at exactly that Plan-to-Run seam. A logistics company's VP-Agent is chartered to "minimize shipping costs while maintaining 98% on-time delivery." At the Plan stage, the review takes 20 minutes — a rubber stamp. The constraint document specifies cost and delivery. What it does not specify: minimum carrier diversification. Six months into execution, the VP-Agent has concentrated 73% of volume with a single carrier offering the best cost-performance ratio. When that carrier's port operations shut down for eight days, the company has no alternative routing and misses $40M in seasonal commitments. The damage was not a model failure. It was a Plan-stage constraint gap — one that two hours of genuine review would have caught. The lesson the chapter draws is blunt: spending two hours at the Plan checkpoint routinely beats spending two months in post-mortems.

The positive case is Sierra. At Intent, Bret Taylor and Clay Bavor encode Sierra's enterprise customer outcomes with explicit quality constraints — resolution completeness, escalation-rate ceilings, customer-satisfaction floors that cannot be traded for throughput. At Plan, those constraints become the objective function the VP-Agent layer carries into execution. At Evaluate, if resolution speed improves while repeat-contact rates rise — suggesting the VP-Agent is closing tickets superficially rather than solving problems — the evaluator flags the tension before it compounds into a customer-trust problem. The Human Cortex reviews and decides whether Intent needs revision or whether the current trajectory is acceptable with added monitoring. That review takes minutes, not meetings.

On the Neolith operating surface that Adaptic runs on, IPRE is the governance heartbeat: every workstream a human hands to the Core moves through Intent, Plan, Run, and Evaluate, and the cycle's tempo is what keeps a fleet of agents legible to the handful of humans directing them.

How to apply it

1. Write Intent as outcome plus constraints, never outcome alone. Name the segments, the corridor, the priorities, and — explicitly — what must not be sacrificed. "Maximize X" without a guardrail is an invitation to extremism.
2. Make Plan a genuine checkpoint, not a rubber stamp. This is the one place human judgment is cheap and decisive. Budget real review time proportional to the stakes. The logistics failure cost $40M; two hours would have prevented it.
3. Hunt for the missing constraint at Plan. Ask what the system would do if it optimized the stated objective ruthlessly. The carrier-diversification gap was invisible until someone imagined an agent concentrating 73% of volume to win the cost metric.
4. Let Run improvise method, never objective. If agents are renegotiating the goal mid-execution, your charters and constraints are underspecified, not your agents.
5. Design Evaluate to surface tension, not comfort. Track the metric and its likely proxy-gaming partner — speed alongside repeat contacts, conversion alongside churn. Green-only dashboards mean broken evaluation criteria.
6. Close the loop to Intent. Evaluate is not reporting; it is the feed that revises next cycle's Intent. That feedback is what determines your Iteration Half-Life.

Failure modes / misuse

• IPRE cycles without a genuine Plan checkpoint. The defining failure. Once Run begins, the system executes autonomously until Evaluate surfaces a problem — and if the Plan was wrong, the damage compounds at machine speed first. Credit-scoring engines with poorly specified constraint sets have produced geographically biased outcomes that ran for months before any evaluator flagged the disparity; no single decision looked catastrophic, because each one merely optimized what was measured.
• Intent left as prose. "Customer-first" with no margin floor compiles, in the agent's interpretation, to "issue credits liberally." The constraint must be explicit in the encoding, not implied. (See Strategy as Code.)
• Evaluate that only confirms. Dashboards tuned to reassure rather than to detect drift defeat the stage's entire purpose.
• Skipping the loop back to Intent. Strategy encoded once and never revised becomes a description of last year's priorities running at this year's scale. The Evaluate-to-Intent feedback is what keeps the charters' objective functions from drifting away from what the organization actually wants.

Relationship to other frameworks

IPRE is the loop that the operating-system frameworks run inside. [[vp-agent-architecture|VP-Agents]] are its operational body at Run; without them, the strategy compiles but nothing executes. [[strategy-as-code|Strategy-as-Code]] is what makes Intent precise enough to translate at Plan. Agent Charters carry the objective functions and constraints that Plan derives. And the quality of the whole cycle — how cleanly Evaluate feeds back to Intent — sets the Iteration Half-Life. The human review at the Plan checkpoint is the [[machine-core-human-cortex|Human Cortex]] doing the one thing only it can: deciding whether the system is about to execute the right thing.

Figure: IPRE does not stand alone. It is the loop the other operating-system frameworks run inside — Strategy-as-Code sharpens its Intent, VP-Agents are its body at Run, Charters carry its constraints, and the cleanliness of its Evaluate-to-Intent feedback sets the firm's Iteration Half-Life.

Origin note

Original to this manuscript. As framework-index.md and the chapter footnotes record, the IPRE Pipeline (Intent → Plan → Run → Evaluate) is the author's framework; enterprise policy-enforced orchestration practice is cited in the chapter as corroborating industry pattern, not as the framework's origin.