Mind the Intent Gap: A Methodology for Collaborative Agent-Driven Software Engineering

1 Introduction

• AI tools did not confirm expected delivery gains: experienced open-source developers took 19% longer with AI, while Cursor’s early speed gains faded as code grew more complex [3] [5].The paper argues that speeding up individual steps does not necessarily speed up the work as a whole.Source: §1 · p. 1
• People in human-designed workflows implicitly filled gaps in documents, carried reasoning between stages, and shared ideas they had worked out alone.The authors hypothesize that these unrecorded contributions help explain why agents may not fit existing processes.Source: §1 · p. 1
• CAIA offers a collaborative team methodology to bring more rigour to software engineering with agents, rather than relying on people to blindly trust agent output.Source: §1 · p. 1
• For each Change, a small team records its Intent; agents evolve the Product from that intent and propose Knowledge Base updates.The Knowledge Base describes the Product a Project maintains, such as its code, tests, and documentation.Source: §1 · p. 1
• The paper presents the challenges, CAIA concepts and Change lifecycle, reports current company use, and outlines a planned student evaluation.Source: §1 · p. 1

2 Related work

• Vibe-coding studies describe a shift toward managing agent context [8], while requirements research traces slippage between those deciding what to build and those building it.Agile values individuals and interactions over processes and tools, but agent conversations can leave colleagues out of the exchange.Source: §2 · p. 1
• Spec-driven development uses structured specifications to drive agents [1], while Krentsel et al. [6] address the intent-to-written-requirement gap with revision loops after the build.Source: §2 · p. 2
• CAIA differs by aligning several disciplines before building, retaining decisions across changes in a Knowledge Base, and governing who drives agents and when intent is frozen.Source: §2 · p. 2
• Building on Storey’s account of cognitive and intent debt [9], CAIA records each change’s intent and reasoning and extends architecture decision records across the product.The authors say they know of no existing methodology combining collective intent alignment, a single agent driver, and a Knowledge Base updated with every change.Source: §2 · p. 2

3 Challenges when agents join the team

• A team’s work on a change involves deciding what it should achieve, deciding how to achieve it, and carrying it out.Source: §3 · p. 2
• The resulting Product comprises the deliverables, components, and documents needed to create and maintain it, in multiple versions.It can range from code to a website and evolves over time because it has value of its own.Source: §3 · p. 2
• A team may work on several Projects, each maintaining its own Product.Source: §3 · p. 2

C1. The intent gap

• The Intent Gap is the difference between a change’s Intent and what its Stated Intent declares.The team first agrees on what it wants, then writes that intent for a colleague or agent.Source: C1. The intent gap · p. 2
• Before agents, developers could use system knowledge, remembered discussions, and colleagues to close gaps left by thin tickets.Source: C1. The intent gap · p. 2
• An agent works from the task’s LLM context and training, and may lack access to the colleague who knows an unanswered point.Source: C1. The intent gap · p. 2
• When information is left unanswered, an agent may invent or ignore it while returning an apparently complete result, leaving the intent gap as a bottleneck.Source: C1. The intent gap · p. 2

• Reasoning behind each deliverable stays with the team that produced it rather than passing to the next team.Source: C2. Reasoning lost between teams · p. 2
• Agents can expand a three-line need into a forty-page document, then compress it into a five-bullet summary.Source: C2. Reasoning lost between teams · p. 2
• When nobody reads the full documents, the original intent is lost between teams.Source: C2. Reasoning lost between teams · p. 2

• Notes, sketches, ideas, and other reasoning often remain undocumented or in individuals’ minds.Source: C3. Private reasoning · p. 2
• Agents keep delegated reasoning private in chat histories and memories maintained about individual users.Source: C3. Private reasoning · p. 2
• Different people can produce divergent versions of a change with their personal agents and distinct LLM contexts.Source: C3. Private reasoning · p. 2
• Merging those versions by hand reintroduces coordination costs.Source: C3. Private reasoning · p. 2

4.1 Requirements

• 4.1 Requirements: CAIA requires explicit intent before agents act, reasoning to travel with each change, and team contributions to one shared stated intent.Source: §4.1 · p. 2

4.2 Concepts

• 4.2 Concepts: A Project evolves its Product one Change at a time and maintains a Knowledge Base of product purpose, organization, conventions, decisions, and reasons.Source: §4.2 · p. 2

4.3 Actors

• 4.3 Actors: A Change Master owns its Stated Intent and operates Build and Knowledge Agents; Leads contribute disciplinary expertise, while Assistant Agents support human thinking.Only the Change Master operates agents that change the Product or Knowledge Base; assistant output enters the Change only when a human writes it into the Stated Intent.Source: §4.3 · p. 3
• 4.3 Actors: Agents receive a Brief assembled from relevant Stated Intent and Knowledge Base content, plus task instructions.Source: §4.3 · p. 3, Figure 1 · p. 3

5 The life of a Change

• 5 The life of a Change: The process has three phases and seven steps: align intent, build and check the Product, then propose Knowledge Base updates.Source: §5 · p. 3

5.1 Alignment

• 5.1 Alignment: The team builds the Stated Intent block by block, preserving reasoning alongside decisions and enabling actors to draw agent Briefs from it.Analysts explore the change by discipline, independent reviewers raise risks and open questions, and the Change Master records the final go or no-go verdict after discussion.Source: §5.1 · pp. 3–4

5.2 Building

• 5.2 Building: The Change Master drives the Build Agent from the aligned brief; open solution choices can be discussed by Leads as the agent proposes them.The micro-team validates the evolved Product, sending small corrections back to Build and new understanding back to Alignment.Source: §5.2 · p. 4

5.3 Consolidation

• 5.3 Consolidation: After validation, the Knowledge Agent proposes Knowledge Base updates from what changed and the decisions and reasons, which humans review before applying.A Change is not finished until the Knowledge Base is updated, keeping it aligned with the Product for the next Change.Source: §5.3 · p. 4

5.4 Rules of the flow

• 5.4 Rules of the flow: Work advances only when each step is complete and Align gives a go verdict, but it may return to any earlier step.Once Build starts, the Stated Intent and micro-team are frozen so the build brief cannot change underneath the agent.Source: §5.4 · p. 4

6 Current Use and Future Plans

• The platform has been used since early 2026 by about ten Ni2 employees, with 19 Changes opened across 4 Projects.Some Changes were still in progress at the time of writing.Source: §6 · p. 4
• In a first subjective evaluation at Ni2, most team members described the approach positively.Source: §6 · p. 4
• A controlled experiment will assign student groups the same software engineering problem, with some following CAIA and others building freely with or without agents.The groups will be compared on performance, such as time per evolution, and accuracy, such as how well results match the intended change.Source: §6 · p. 4
• The authors hypothesize that CAIA groups will deliver results closer to the intended change, with fewer corrections, in comparable time.Source: §6 · p. 4

7 Conclusion

• CAIA addresses processes designed for people, who filled gaps in written intent, carried reasoning between steps, and shared ideas that had remained private.Source: §7 · p. 4
• For each change, a micro-team aligns on intent before an AI agent builds, and the Knowledge Base retains what the change leaves behind for the next one.Source: §7 · p. 4
• CAIA emerged from an industrial need and is in use at Ni2, with a controlled experiment planned at the University of Mons to measure its effect.Source: §7 · p. 4