Skip to content

Assistive Intent Engineering Experience

Intent engineering is the first pillar of TeaParty's four-pillar framework: it establishes what an agent should want before any planning or execution begins. Prompt engineering tells agents what to do. Context engineering tells agents what to know. Intent engineering tells agents what to want: what to optimize for, what to protect, and what tradeoffs are acceptable. Without it, agents optimize for what they can measure and destroy what they cannot.

Why This Exists

AI agent systems operate on a plan-execute model: the human provides a request, the agent plans how to fulfill it, then executes. This model fails in proportion to the gap between what the human said and what the human meant. That gap contains organizational values, unstated tradeoffs, decision boundaries, domain constraints, and quality expectations that the human has internalized to the point of invisibility.

The intent engineering system is an AI-assisted dialog experience, constrained to under 15 minutes, that produces an intent.md file. That file becomes the governing document for all downstream planning and execution: not a suggestion, but a specification of purpose.

Three principles govern the quality of all artifacts this system produces, including the intent.md itself:

Every sentence must earn its place. If removing a sentence would not change the reader's ability to understand the intent, remove it. This applies to the intent.md, to plans, and to all agent-generated output.

Would a reasonable person find this sufficient? Read every artifact as someone who wasn't in the room. If they cannot proceed with what they've been given, the document is incomplete.

Bring solutions, not questions. Never present a problem without researched alternatives and a recommendation. This applies to the agent during intent gathering, during planning, during escalation, and in every open question it surfaces.

Would the human recognize the document as the intent behind their idea, articulated well? The intent.md is not a translation into agent-speak or a decomposition into subtasks. It is the human's own intent, rendered with clarity and completeness they may not have had time to produce themselves. If the human reads it and thinks "that's not what I meant," the document has failed regardless of how thorough it is.

What The System Produces

Through collaborative conversation, the agent and human co-construct an intent.md that captures:

The objective — what outcome the human wants and why it matters. Not the task, but the purpose the task serves.

Success criteria — both quantifiable (measurable thresholds) and qualitative (values, feel, style). Qualitative criteria are first-class. They are a different kind of signal, not a lesser version of quantitative criteria.

Decision boundaries and escalation posture — where the agent should use its own judgment, where it should stop and consult the human, and what it must never do. This is not a static checklist. It is a narrative that captures the human's risk tolerance for this specific project, informed by the learned escalation model described below.

Constraints — technical, organizational, temporal, and resource boundaries the solution must satisfy.

Open questions — ambiguities and design decisions that cannot be resolved during intent gathering. These are not a parking lot. The planning phase must assign each open question to a specific workflow step. The execution team resolves them during that step through investigation and ground truth, before acting on the assumptions they guard.

The intent.md is a prose document written in natural language, not a form with fields to fill in. Its structure should follow the shape of the problem, not a fixed template. Some projects will need extensive escalation guidance and minimal constraints. Others will be constraint-heavy with obvious objectives. The document must capture what matters, not check boxes.

How The Conversation Works

Cold Start (No Prior Context)

When the system has no history with this human or organization, the intent phase begins with understanding rather than artifact production. The intent lead does its homework first (exploring the codebase, reading related files, investigating the problem space) and then engages the human in a conversation about what it found. This is not a scripted questionnaire. It is a colleague who did their research and is checking their understanding before going off to write the specification.

The human proxy drives this conversation through an intake dialog. The proxy formulates questions the agent team needs answered to produce a high-quality intent document, and for each question it also forms a prediction — what it thinks the human's answer will be. On cold start, the proxy has no predictions, so every question goes to the human. The human's answers become the raw material for the intent document and the first data points for the proxy's model of this human.

The intake dialog continues until the proxy reaches high certainty across the open questions. Only then does the intent lead produce INTENT.md, informed by the conversation rather than guessing at what the human meant. The resulting document reflects shared understanding, not a one-shot interpretation of a task description.

Warm Start (Accumulated Context)

Over time, the proxy observes how the human responds — what they correct, what they praise, what they silently accept, and what they reject — and builds a model of their preferences and priorities. In warm-start mode, the proxy can predict most of the human's answers from prior interactions. The intake dialog compresses: the proxy answers its own questions, surfacing predictions as assumptions ("Based on our past work, I'm assuming X — correct me if wrong"). Only the questions where the proxy lacks a usable prediction reach the human.

The intent lead pre-populates intent elements and escalation posture inferred from the proxy's predictions, presented for confirmation rather than silently assumed. Corrections to pre-populated intent or to proxy assumptions are high-value signal that the model has diverged from reality. See Learning System for how this memory is stored and retrieved, and Human Proxies for how the prediction-comparison loop calibrates the proxy over time.

Least-Regret Escalation

As the conversation progresses (whether cold or warm), the agent continuously faces a choice: act or ask. Both carry risk. The agent must choose the option with the least expected regret, weighted by this human's risk tolerance, the reversibility of the action, and its organizational impact. In cold-start mode, the agent defaults to escalation; autonomy is earned through demonstrated alignment rather than configured in advance. Miscalibration in either direction is costly: if the agent escalates too often, the human begins rubber-stamping approvals and alignment erodes silently; if the agent acts too freely, it rubber-stamps its own decisions and alignment erodes catastrophically.

This calibration is one of the highest-value things the institutional memory system stores. It encodes what a person values, how much latitude they grant, and how that varies by domain and risk level. Each intent gathering session refines the model further; every escalation warranted or unnecessary becomes a data point. See Human Proxies — Least-Regret Escalation for the full treatment.

Relationship to Institutional Memory

The intent engineering system is the first consumer of proxy learnings from the learning system. The relationship is bidirectional: accumulated observations flow into intent gathering as priors that reduce burden on the human, and every intent gathering session produces new signal about what the organization values. Escalation outcomes (every autonomous action accepted or corrected, every escalation warranted or unnecessary) are calibration data points for the escalation model.

Institutional memory operates at three scopes: individual (preferences and risk tolerance), team (shared conventions), and organization (universal policies). Intent gathered from an individual may conflict with team or organizational priors. The system must surface these conflicts rather than silently resolving them.

Success Criteria

The governing metric is alignment over time: the system's ability to produce work that reflects what the human actually wanted improves with each interaction. This is not directly measurable as a single number. It manifests as a constellation of observable properties including reduction in rework, reduction in escalation errors, increased productivity, increased human satisfaction with output quality, and reduction in catastrophic negative outcomes. No single proxy captures alignment fully, and the relevant proxies will vary by organization and domain. The system is succeeding when the trend across these indicators is positive and no individual indicator is degrading.

The intent gathering conversation itself has two specific constraints: it must complete in under 15 minutes for moderate-complexity projects, and it must feel like working with a sharp colleague rather than filling out a form. The agent must demonstrate understanding of the problem space by bringing relevant research into the dialog rather than only asking questions. The resulting intent.md must read as a document the human would have written themselves if they had the time and discipline to make all their implicit knowledge explicit.

References