<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[JuegoTech Engineering]]></title><description><![CDATA[JuegoTech Engineering]]></description><link>https://juegotech.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Sat, 21 Nov 2026 17:10:27 GMT</lastBuildDate><atom:link href="https://juegotech.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Greenfield vs brownfield: how field engineers land a GenAI platform on either]]></title><description><![CDATA[A blank repository and a fifteen-year-old codebase look like opposite problems. The work that makes GenAI shippable on both is surprisingly similar — only…]]></description><link>https://juegotech.hashnode.dev/greenfield-vs-brownfield-how-field-engineers-land-a-genai-platform-on-either</link><guid isPermaLink="true">https://juegotech.hashnode.dev/greenfield-vs-brownfield-how-field-engineers-land-a-genai-platform-on-either</guid><category><![CDATA[generative ai]]></category><category><![CDATA[software development]]></category><category><![CDATA[legacy modernization]]></category><category><![CDATA[AI]]></category><category><![CDATA[Software Engineering]]></category><dc:creator><![CDATA[JuegoTech]]></dc:creator><pubDate>Fri, 20 Nov 2026 21:23:05 GMT</pubDate><content:encoded><![CDATA[<p>Ask a room of engineering leads where generative AI helps most and you will usually hear two confident, opposite answers. One camp says greenfield: no legacy, no tangled dependencies, just let the agents build. The other camp says brownfield: the real payoff is in understanding and modernizing the systems nobody wants to touch.</p>
<p>Both camps are partly right, and both tend to skip the same thing. Whether the repository is empty or enormous, generated code only becomes shipped software when someone decides what the work is for, scopes what the agents may touch, and reviews what they produce before it merges. At JuegoTech, that "someone" is a pairing of your leads with our forward deployed engineers (FDEs), and the scaffolding they work inside is a platform rather than a chat window.</p>
<p>This post walks through how that landing differs between a greenfield build and a brownfield system, and where it deliberately does not differ at all.</p>
<h2>The shared skeleton</h2>
<p>Before the differences, the constant. Every engagement we describe publicly follows one cadence: <strong>Align → Prove → Ship</strong>.</p>
<ul>
<li><strong>Align</strong> (roughly weeks 0–2) is about outcomes, constraints and non-goals. Stakeholder workshops, context packs, bring-your-own-key model setup, and agreement on who reviews what before any build step is unlocked.</li>
<li><strong>Prove</strong> (roughly weeks 2–6) is a thin vertical slice, built under planning and debugging modes, that ends in a real decision: kill, pivot or continue, based on evidence.</li>
<li><strong>Ship</strong> is ongoing: a production minimum lovable product delivered through develop, review and fix loops, with agent review before every generative merge and practice transferred to your team.</li>
</ul>
<p>That skeleton is the same on day one of a blank repo and day one of a legacy monolith. What changes is what fills the first phase, and where the risk sits.</p>
<h2>Greenfield: the risk is that intent drifts</h2>
<p>On a greenfield build, there is no existing code to break. That is exactly why it feels safe to let agents run fast, and exactly why it goes wrong. With nothing pushing back, a vague brief turns into a large volume of plausible code that nobody quite asked for.</p>
<p>So on greenfield work the FDE's early energy goes into locking intent:</p>
<ol>
<li><strong>A brief that can be argued with.</strong> Outcomes, constraints and explicit non-goals, written down in Align, so that "build a dashboard" becomes a scoped slice with success signals someone can check.</li>
<li><strong>Plan before Build.</strong> A structured plan is drafted (models can help draft it) and a human approves it before any tool is allowed to make side effects. On a blank repo, this is the only brake there is.</li>
<li><strong>Craft roles named early.</strong> Architecture, UX and release ownership are assigned on the squad, so design decisions do not get made by default inside a generated diff.</li>
<li><strong>A first slice that is genuinely thin.</strong> The goal of Prove is not to impress anyone with volume. It is to put one end-to-end path in front of stakeholders and decide, with evidence, whether to keep going.</li>
</ol>
<p>Greenfield is where JuegoTech's bespoke-product work sits: embedded operators and your owners sharing one workspace from brief through minimum lovable product. The temptation there is speed. The discipline is making sure the speed is pointed at the agreed outcome.</p>
<h2>Brownfield: the risk is that nobody understood what was already there</h2>
<p>Brownfield is the opposite failure mode. The intent is usually clear ("modernize this", "add this capability to that system"), but the existing code holds years of decisions, undocumented rules and edge cases. An agent that writes confidently into a system it has not mapped is a liability, however good the model is.</p>
<p>So on brownfield work the FDE's early energy goes into understanding before writing:</p>
<ol>
<li><strong>Read-safe mapping first.</strong> Work starts in an ask/explore mode that can search, read and research the repository but cannot write to it. The point is to turn the existing code into context before anyone proposes a change.</li>
<li><strong>Context packs built from the real system.</strong> Repo slices, domain documents and success signals are assembled deliberately, rather than pasted ad hoc into a chat window.</li>
<li><strong>Path grants that match the risk.</strong> Agents are scoped to the parts of the system they are allowed to read and execute against. In regulated settings such as hospital operations or financial services, this is what lets GenAI into the repository at all.</li>
<li><strong>Strangler slices, not a big-bang rewrite.</strong> The first Prove slice replaces or extends one bounded piece of behaviour, with review and acceptance evidence, while the rest of the system keeps running.</li>
</ol>
<p>The healthcare use case on our site is labelled as an illustrative composite, not a named client, and it describes exactly this shape: mapping with ask/explore, PHI-aware path grants, strangler slices and agent review, so that GenAI speeds delivery without bypassing compliance.</p>
<h2>Side by side</h2>
<table>
<thead>
<tr>
<th></th>
<th>Greenfield</th>
<th>Brownfield</th>
</tr>
</thead>
<tbody><tr>
<td>Main risk</td>
<td>Intent drifts; volume without direction</td>
<td>Changes break what nobody mapped</td>
</tr>
<tr>
<td>First move in Align</td>
<td>Lock the brief, non-goals and success signals</td>
<td>Read-safe mapping of the existing system</td>
</tr>
<tr>
<td>What the Prove slice tests</td>
<td>Is this the right product shape?</td>
<td>Can we change this safely, one slice at a time?</td>
</tr>
<tr>
<td>Where path grants matter most</td>
<td>Keeping agents inside the agreed scope</td>
<td>Keeping agents away from sensitive or fragile areas</td>
</tr>
<tr>
<td>What "done" protects against</td>
<td>Shipping something nobody asked for</td>
<td>Shipping a regression into a live system</td>
</tr>
</tbody></table>
<h2>What does not change at all</h2>
<p>Here is the part that surprises teams: from the end of Prove onwards, the two engagements look almost identical.</p>
<p>The same definition of done applies to both. Before a slice is called shippable, the plan for it is approved, review rules and the craft role are clear, agent review has passed or its concerns have been dispositioned, the context pack and path grants are still accurate for that cut, the story or epic status is updated, and acceptance evidence (tests, a demo or review notes) is attached.</p>
<p>The same ceremonies apply too. Standups keep parallel agent threads visible, including what is blocked and what needs a human. Plan reviews happen before build unlocks side effects. Diff reviews end in pass, fix or escalate, never a silent merge.</p>
<p>And the same exit applies. Whether the code was born last month or ten years ago, success means two things: the mission outcome, and practice left with your team. The signals we look for are the same in both cases. Rework after plan approval trends down. Your own engineers can run plan, build and review without an FDE watching over them. Dispositioning review findings before merge becomes the default rather than the exception.</p>
<h2>What your team brings, in either case</h2>
<p>Field engineering is not a handover. On both greenfield and brownfield missions, the client side provides:</p>
<ul>
<li>product and engineering decision-makers for the gate ceremonies,</li>
<li>repository access, domain documents and security-approved model keys,</li>
<li>clear constraints on data locality and what agents may touch,</li>
<li>and a real willingness to treat planning and review as ship gates, not paperwork.</li>
</ul>
<p>That last point matters more than the repo's age. A blank repository with no gates produces unreviewed output quickly. A legacy system with real gates can be modernized one safe slice at a time.</p>
<h2>Choosing your starting point</h2>
<p>If you are deciding where to point generative AI first, the more useful question is not "greenfield or brownfield?" but "where is our risk?" If the risk is building the wrong thing, spend Align on intent. If the risk is breaking the thing that already pays the bills, spend Align on mapping. Either way, Prove should end in an honest decision, and Ship should leave your team owning the cadence.</p>
<p>We have written up the full engagement cadence, the definition of done and the different engagement shapes (embedded squad, fixed-phase MLP, and enablement or co-delivery) on our <a href="https://www.juegotech.com/delivery">delivery page</a>. If you have a system in mind, blank or brownfield, that is a good place to start the conversation.</p>
<p><em>JuegoTech builds GenAI-native software platforms for enterprises and deploys forward deployed engineers to land them in the field. We are based in Bengaluru.</em></p>
]]></content:encoded></item><item><title><![CDATA[Companion admin, seats, and the Agents Window: how a field squad stands up]]></title><description><![CDATA[Before any agent writes a line of code, someone has to decide who is in the room, which keys are in play, and where the ceremonies happen.]]></description><link>https://juegotech.hashnode.dev/companion-admin-seats-and-the-agents-window-how-a-field-squad-stands-up</link><guid isPermaLink="true">https://juegotech.hashnode.dev/companion-admin-seats-and-the-agents-window-how-a-field-squad-stands-up</guid><category><![CDATA[genai]]></category><dc:creator><![CDATA[JuegoTech]]></dc:creator><pubDate>Fri, 20 Nov 2026 09:37:09 GMT</pubDate><content:encoded><![CDATA[<p>Most writing about GenAI in software delivery starts at the exciting part: the agent that drafts a feature, the model that explains a legacy module, the diff that appears in seconds. Very little of it talks about the first week of an engagement, when a squad of engineers is trying to get set up inside an institution that has security reviews, procurement rules, and a repository nobody outside the company has touched.</p>
<p>That first week decides a lot. If access is improvised, every later step inherits the improvisation. If seats are shared, history stops being attributable. If ceremonies happen in scattered chat threads, nobody can reconstruct why a plan was approved. This post walks through how we think about standing up a field squad on a GenAI-native platform, using the pieces JuegoTech describes on its platform: provisioned seats, companion admin, BYOK models, and the Agents Window.</p>
<h2>Start with ownership, not tooling</h2>
<p>The most important decision is who owns the accounts. In our model, the client organisation owns the seats. Accounts are invite-only, devices are approved, and the org's admin controls who is in. Forward Deployment Engineers (FDEs) join as collaborators inside that org rather than bringing their own workspace and asking the client to trust it.</p>
<p>This sounds like a small administrative detail, but it changes the shape of the engagement:</p>
<ul>
<li><strong>Exit is clean.</strong> When the embed steps back, the client does not need to migrate anything. The seats, history, and configuration were theirs all along.</li>
<li><strong>Access reviews are normal.</strong> Security teams can review collaborator access the same way they review any other vendor access, because it lives in a system they administer.</li>
<li><strong>Nothing is trapped.</strong> Work is not sitting in a vendor sandbox the client cannot audit.</li>
</ul>
<p>A practical checklist for day one:</p>
<ol>
<li>The client admin creates the org and invites the client leads first.</li>
<li>Device approval is switched on before any FDE is invited.</li>
<li>FDEs are invited as collaborators, one seat per person, never shared.</li>
<li>The admin confirms who can approve plans and who can disposition review findings.</li>
</ol>
<h2>Keys under the client's procurement</h2>
<p>The second decision is which models the squad may use and whose keys pay for them. JuegoTech's platform is built around multi-model BYOK: the client brings approved keys for Anthropic Claude, OpenAI, Google Gemini, or private endpoints, under its own procurement.</p>
<p>Two things follow from this. First, the model choice is a client policy decision, not a vendor default. A regulated team can restrict the squad to the providers its legal and security teams have already cleared. Second, usage stays attributable in the companion. When finance asks what generative work cost and who ran it, the answer comes from the client's own records instead of a vendor invoice.</p>
<p>In practice we suggest the admin records, before the first Build session:</p>
<ul>
<li>Which providers are approved and for which repositories.</li>
<li>Whether any private endpoint must be used for sensitive code paths.</li>
<li>Who reviews usage in the companion, and how often.</li>
</ul>
<h2>Scope before speed: path grants and run policy</h2>
<p>Once people and keys are in place, the next question is what agents are allowed to touch. Path grants and run policy scope what agents may read and execute. This is where the brownfield reality of most enterprise work shows up. A large repository usually contains areas that are fine to explore, areas that are fine to change under review, and areas that nobody outside a small group should be editing at all.</p>
<p>Setting path grants early prevents the most common failure of GenAI tooling in big codebases: an agent with broad access making a confident change somewhere it should never have been. It also makes the first sessions calmer. Engineers can explore freely in read-safe Ask mode, knowing the boundaries are already drawn.</p>
<h2>The Agents Window as the room where decisions happen</h2>
<p>With the foundations in place, the squad needs a shared place to work. On the JuegoTech platform that is the Agents Window, which hosts ceremonies and parallel threads alongside the desktop IDE, the terminal CLI, and the web companion.</p>
<p>We treat the Agents Window as the room where decisions are made visible. A typical early rhythm looks like this:</p>
<ul>
<li><strong>Mapping threads.</strong> FDEs run read-safe exploration against legacy or blank repositories, and the findings are discussed in the open rather than in private notes.</li>
<li><strong>Plan sessions.</strong> Scope is researched and clarified, then locked into a structured plan. Humans approve the plan before Build mode unlocks side effects.</li>
<li><strong>Parallel threads.</strong> Separate slices can move at the same time, each with its own plan and its own review, without blurring into one long conversation.</li>
<li><strong>Review dispositions.</strong> Agent Review scores the diff and produces findings with a verdict before anything merges. FDEs and client leads decide together whether to pass, fix, or escalate.</li>
</ul>
<p>The point is not ceremony for its own sake. It is that every generative change has a visible trail: who mapped the area, who approved the plan, which model ran, and who accepted the review. There are no silent generative merges.</p>
<h2>What the client leads actually do</h2>
<p>A field squad is not only the FDEs. Client leads are part of it from the start, and the setup should make their role explicit. In our experience, leads spend the first week on three things:</p>
<ol>
<li><strong>Approving plans.</strong> They are the people who know which constraints matter, so their approval is what unlocks Build.</li>
<li><strong>Reading review findings.</strong> They see the same Agent Review output as the FDEs and share the decision on each finding.</li>
<li><strong>Watching usage and history.</strong> The companion gives them a view of who is running what, which keeps the engagement grounded in facts rather than status updates.</li>
</ol>
<p>When leads are involved this way, the transfer of practice at the end of an embed is much easier. They have already been running the cadence; the FDEs are simply stepping back.</p>
<h2>A simple stand-up sequence</h2>
<p>Putting it together, the order we recommend for standing up a squad is:</p>
<ol>
<li>Org and admin created by the client; leads invited.</li>
<li>Device approval on; FDEs invited as collaborators.</li>
<li>Approved BYOK providers configured; usage owner named.</li>
<li>Path grants and run policy set for the target repositories.</li>
<li>First mapping threads in Ask mode, discussed in the Agents Window.</li>
<li>First structured plan approved by a client lead; Build unlocks.</li>
<li>First Agent Review dispositioned jointly before merge.</li>
</ol>
<p>None of these steps is glamorous, and none of them should take long. But skipping them is how GenAI pilots end up producing impressive demos that never reach production, because nobody can explain to a risk committee who did what and why.</p>
<h2>Closing thought</h2>
<p>The model is the engine, but the setup is the line it runs on. Seats owned by the client, keys under its procurement, scoped access, and a shared room for ceremonies are what let institutions say yes to GenAI in their repositories. If you want to see how these foundations fit together with modes, plans, and Agent Review, the <a href="https://www.juegotech.com/platforms">JuegoTech platforms overview</a> lays out the full Map to Operate line.</p>
<p><em>JuegoTech builds GenAI-native software platforms for enterprises and deploys them in the field with Forward Deployment Engineers, from Bengaluru.</em></p>
]]></content:encoded></item></channel></rss>