---
title: "Vlad Zoff (@vladzoff) | Hashnode"
description: "Building Codapult and Codapult Guard. Read the latest articles by Vlad Zoff on Hashnode."
url: "https://hashnode.com/@vladzoff"
---

@vladzoff

# Vlad Zoff

@vladzoff · Joined July 2026

Building Codapult and Codapult Guard

[Website](https://codapult.dev/)

## Comments

[The replay idea is useful, but full traces can become another copy of production data. Tool inputs and outputs may contain API tokens, customer records, emails or other sensitive values. I'd want field-level redaction before events are persisted, not just a warning to be careful when exporting them. Do you have a way to configure that, or is it currently up to the application to sanitize the payloads?](https://hashnode.com/posts/how-do-i-debug-an-ai-agent-run-after-it-finished/6aea60ce421b4fe5e077c984/comment/6aec23745c230ec8005922e0)

Comment · Article · Nov 5 · [How Do I Debug an AI Agent Run After It Finished?](https://hashnode.com/posts/how-do-i-debug-an-ai-agent-run-after-it-finished/6aea60ce421b4fe5e077c984)

[I'd add validity metadata to promoted knowledge: source or commit, scope, owner, and what superseded it. An architectural decision can be durable but apply only to one module or version. Without scope and freshness, project memory becomes another stale source of truth. Agents should revalidate older notes against the current code and tests before treating them as constraints.](https://hashnode.com/posts/session-memory-is-not-project-memory-it-fixes-the-same-complaint/6ab3ae06ab5703ad22d98586/comment/6ae833e119be44d2c35fb36b)

Comment · Article · Nov 2 · [Session memory is not project memory. It fixes the same complaint.](https://hashnode.com/posts/session-memory-is-not-project-memory-it-fixes-the-same-complaint/6ab3ae06ab5703ad22d98586)

[I'd also carry those constraints through to the output, not stop at retrieval. An agent may be allowed to read two sources separately but not combine them into a summary for a recipient who can only access one of them. So the policy needs to cover both source access and destination constraints, including derived content. Otherwise narrow retrieval can still result in a broader disclosure.](https://hashnode.com/posts/what-would-it-take-to-make-agent-context-useful-without-making-access-broad/6aba70e458dc0108a22c53e0/comment/6ae833d46ead3872ff3a7460)

Comment · Article · Nov 2 · [What would it take to make agent context useful without making access broad?](https://hashnode.com/posts/what-would-it-take-to-make-agent-context-useful-without-making-access-broad/6aba70e458dc0108a22c53e0)

[I try to catch them as early as possible, ideally during planning, rather than relying on review to uncover them. I ask the agent to separate confirmed requirements from assumptions and open decisions. If a decision affects user-visible behavior, data contracts, security, or scope, it should stop and ask instead of silently choosing. For low-risk, reversible details, it can make a reasonable choice, but should record the assumption. Review is still important, but it should validate those decisions rather than be the first place they become visible.](https://hashnode.com/posts/how-much-specification-does-an-agent-need/6aa46924a9b5f506090a6413/comment/6abdf3b252950c0a4520735f#reply-6ae73c915cc46e26e58aa7fc)

Reply · Article · Nov 1 · 1 · [How much specification does an agent need?](https://hashnode.com/posts/how-much-specification-does-an-agent-need/6aa46924a9b5f506090a6413)

[The boundary I'd keep explicit is that a skill can explain how to use a tool, but it can't be the only place where a security rule lives. The tool or server still needs to validate inputs and enforce the current user's permissions. That way the instruction helps the model, while the execution boundary protects the system even when the model ignores the guidance.](https://hashnode.com/posts/tools-mcp-and-skills-what-each-one-actually-does/6a609b89acfc52517f3ce85a/comment/6ae20a4a46d69ef92fd415e2)

Comment · Article · Oct 28 · [Tools, MCP, and Skills: What Each One Actually Does](https://hashnode.com/posts/tools-mcp-and-skills-what-each-one-actually-does/6a609b89acfc52517f3ce85a)

[Token reduction is a good start, but I'd also test retrieval quality after a symbol is renamed or moved across modules. An index that returns fewer tokens but points to stale locations can make the agent confidently edit the wrong implementation. Measuring recall on a small set of real repo questions would help show whether the savings are actually useful.](https://hashnode.com/posts/ai-agent-codebase-indexer-skill-spelunk/6a609b9ba874e83d3cbc81e5/comment/6ae20a3277d5cbaa0c48c1bc)

Comment · Article · Oct 28 · [My AI coding agent kept grepping my repo to find one function — so I cut its token usage by 90%](https://hashnode.com/posts/ai-agent-codebase-indexer-skill-spelunk/6a609b9ba874e83d3cbc81e5)

[A shared file helps only where the rules are genuinely shared. I'd keep one canonical source for common constraints and generate thin tool-specific files where the tools differ, then have CI detect drift. Trying to force every tool's behaviour into one universal file can just move the sprawl into conditionals and exceptions.](https://hashnode.com/posts/agents-md-was-supposed-to-end-the-sprawl-i-counted-the-files-on-one-laptop/6a824fd256459d42c0fa4d84/comment/6adf60b1fe24eeffc1f92873)

Comment · Article · Oct 26 · [AGENTS.md was supposed to end the sprawl. I counted the files on one laptop.](https://hashnode.com/posts/agents-md-was-supposed-to-end-the-sprawl-i-counted-the-files-on-one-laptop/6a824fd256459d42c0fa4d84)

[For an agentic workload, I'd compare cost per successfully completed task rather than token price alone. Tool round trips, retries, schema failures and repair attempts can erase a large input-price advantage. A small representative eval set with fixed acceptance criteria gives a more useful comparison than running each model on a single prompt and comparing the token bill.](https://hashnode.com/posts/gemini-4-argon-vs-gpt-6-astra-vs-claude-api-pricing-and-access-for-developers-october-2026/6acb160b911efc43bbba9469/comment/6ad21c392f8685a3137a29be)

Comment · Article · Oct 16 · [Gemini 4 Argon vs GPT-6 Astra vs Claude: API Pricing and Access for Developers (October 2026)](https://hashnode.com/posts/gemini-4-argon-vs-gpt-6-astra-vs-claude-api-pricing-and-access-for-developers-october-2026/6acb160b911efc43bbba9469)

[Typed output validation solves format drift, but a valid schema doesn't guarantee that the agent chose the right action. I'd evaluate routing decisions separately from execution outcomes, and treat validation failure as a hard stop before side effects rather than silently repairing the object and continuing. Otherwise the pipeline can be perfectly type-safe while still taking the wrong action.](https://hashnode.com/posts/typesafe-jev-the-870m-shift-from-all-in-one-llms-to-decision-assembly-lines/6aca09e58befd0341e1e98f3/comment/6ad21c24b280e06fab44cf6d)

Comment · Article · Oct 16 · [TypeSafe Jev: The $870M Shift from All-in-One LLMs to Decision Assembly Lines](https://hashnode.com/posts/typesafe-jev-the-870m-shift-from-all-in-one-llms-to-decision-assembly-lines/6aca09e58befd0341e1e98f3)

[For agent workflows, I'd document persistence as a lifecycle matrix: process state, working directory, durable volume, database state and generated artifacts. They can have different retention and reset rules even when the provider markets the environment as persistent. A small test that restarts the sandbox and checks which artifacts survive is more useful than treating persistence as a single yes/no property.](https://hashnode.com/posts/free-compute-october-2026/6abe4f2998e3cf309fd461a7/comment/6ad21c09402dd7b98c9766c0)

Comment · Article · Oct 16 · [Free compute in October 2026: new backends and temporary VMs](https://hashnode.com/posts/free-compute-october-2026/6abe4f2998e3cf309fd461a7)
