AI Coding Needs Human Control, Not Speed: The AEGIS Trust Approach

AI coding assistants can significantly boost productivity, but developer ProlixoTech Inc. discovered a commonly overlooked pain point during long-term use: getting AI to complete a task and getting AI to stay aligned with intent are two entirely different things. To address this, he developed the open-source VS Code extension AEGIS Trust, aiming to re-embed human control into the AI-assisted development workflow.

When AI Assistants Drift From Your Intent

In complex project development, AI assistants frequently exhibit five types of behavioral drift: expanding scope without being asked, simplifying tasks that should be delivered in full, repeating work already completed, skipping explicitly specified requirement steps, and making changes based on their own understanding rather than actual instructions.

The more vexing problems emerge during long development sessions. AI loses earlier decisions, forgets important instructions, or misremembers what has already been done. ProlixoTech admits in the original post: "I sometimes find that I spend far more time than expected checking, correcting, and re-steering AI-generated work."

This reveals a key fact: the bottleneck in AI coding isn't code generation speed—it's the continuity of intent alignment.

Intent Alignment here refers to an AI assistant's ability to accurately understand and follow a developer's true intent across sustained interactions, as distinct from single-task completion quality. In long-session scenarios, as the context window fills with large amounts of code, logs, and intermediate results, early explicit constraint instructions are easily diluted or overwritten, causing the AI to make decisions based on local context rather than global goals. This is a general limitation of current large language models in long-context reasoning, not a defect specific to any particular tool. Therefore, the continuity of intent alignment requires external mechanisms (such as explicit records, periodic reviews, and behavioral boundary constraints) to compensate for the unreliability of the model's own memory and reasoning.

Core Thesis: AI Should Assist, Not Take Over

Based on these experiences, the author draws a conclusion: AI should assist developers, not take over the development process. The real challenge isn't getting AI to write better code—it's maintaining human control, clear boundaries, accountability, and the ability to recover after mistakes.

This stance contrasts with the current trend among AI coding tools to emphasize "automation" and "end-to-end generation." AEGIS Trust's starting point isn't raising the ceiling of AI capability but setting a floor for AI behavior—ensuring developers always retain the final say over "what to do, what to change, and what to execute."

The Four Core Mechanisms of AEGIS Trust

According to the project description, AEGIS Trust version 0.1.1 is a free, open-source VS Code extension developed by ProlixoTech Inc., currently featuring four capabilities:

Local File Baselines and Snapshots

For selected important project files, it supports establishing a known starting point and performing recovery-oriented operations. This is essentially laying an escape route before AI starts working, reducing the risk of irreversible changes.

File Baseline and Snapshot are foundational concepts in version control: a baseline refers to a starting state explicitly marked as "known good," against which subsequent changes are compared; a snapshot is a complete record of file contents at a given moment. The value of combining the two lies in providing rollback capability—when AI modifications deviate from expectations or introduce hard-to-locate problems, developers can quickly restore to the snapshot point without relying on manual line-by-line undo. Unlike full version control systems such as Git, such mechanisms typically operate only on selected files at a lighter granularity, aiming to provide low-cost recovery points during short cycles of high-frequency AI modifications rather than replacing formal version management workflows.

AI-Suggested Command Review

Terminal commands suggested by AI are not executed immediately but first submitted to the AEGIS review process. Developers get an extra opportunity to inspect the proposed action before deciding whether to proceed.

Reviewed Terminal Handoff

Supported Protected Terminal workflows evaluate proposed commands before handing allowed text off to the terminal. The key point: commands are not executed automatically—the developer retains the decision to run them.

Terminal Handoff refers to the process by which an AI assistant hands off generated commands to a real terminal environment for execution. In default workflows without review, AI-proposed commands may be executed automatically, creating risks of destructive operations (such as accidentally deleting files, overwriting configurations, or running irreversible scripts). "Reviewed handoff" means inserting an evaluation and confirmation step before commands enter the terminal, preserving the human's veto power before commands actually run. This pattern aligns with the "Human-in-the-loop" design principle, and its core trade-off is sacrificing a small amount of automation fluidity in exchange for controllability over high-risk actions.

Local Work Receipts

Local Markdown records are generated from AEGIS-managed activity to provide reference for reviewing supported operations. The author explicitly states that these receipts are not a complete record of all AI assistant behavior.

Facing Limitations Transparently: What It Can't Do Yet

The author provides a clear explanation of AEGIS Trust's capability boundaries—something commendable in open-source tools:

• It does not currently guarantee that the AI assistant correctly remembers prior instructions, nor that it stays within the original task scope;
• It does not automatically intercept every action performed by external AI agents, ordinary terminals, or other software;
• File protection and review capabilities apply only to specific supported workflows;
• It is not a general-purpose sandbox and cannot replace independent backups.

"AEGIS Trust is a pragmatic starting point, not a claim that the problem is fully solved." This candor about limitations actually makes the tool's positioning more credible.

General-purpose Sandbox is an isolated execution environment that restricts file system, network, and process permissions so that programs running inside cannot affect the host system even if they behave abnormally. AEGIS Trust explicitly states it is not a general-purpose sandbox, meaning it does not provide system-level isolation guarantees—it only reviews and snapshots files and commands within specific supported workflows. Understanding this distinction is important: if you need to guard against privilege escalation from code or AI agents of arbitrary origin, you still need to rely on independent mechanisms such as OS-level sandboxes, containers, or virtual machines. Review tools and isolation tools solve problems at different layers.

Redefining the Metrics for AI Coding

The principle behind the project is worth the entire industry's consideration: the future of AI-assisted development should not be measured solely by how much code AI can generate or how quickly it completes tasks, but also by how reliably developers can guide, understand, review, and control that work.

AI can propose, can assist, can accelerate—but developers should always control what gets done, what gets changed, and what gets executed. At a time when AI coding tools are racing toward greater automation, AEGIS Trust reminds us: control itself is an underappreciated form of productivity.

AEGIS Trust 0.1.1 is available as a free, open-source release on the VS Code Marketplace and GitHub, and the author is especially eager to hear feedback from developers who have experienced similar problems.

One-line Summary

An open-source VS Code extension, AEGIS Trust, seeks to re-embed human control into the AI coding workflow.

Paragraph Summary

Developer ProlixoTech Inc. found through long-term use of AI coding assistants that getting AI to complete tasks and keeping AI aligned with intent are two different things: AI often expands or shrinks scope, repeats work, skips steps, and forgets instructions, and the cost of checking and correcting can offset the productivity gains. He therefore developed the free, open-source VS Code extension AEGIS Trust 0.1.1, which provides four mechanisms—local file baselines and snapshots, AI-suggested command review, reviewed terminal handoff, and local work receipts—with the core idea that commands are not executed automatically and developers retain final decision-making authority. The author frankly acknowledges the tool's limitations: it doesn't guarantee correct AI memory, doesn't intercept all external actions, and isn't a general-purpose sandbox. His thesis is that AI should assist rather than take over development, and that metrics should include whether developers can reliably guide, review, and control AI work.