<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Tiru Kavala</title>
    <description>The latest articles on DEV Community by Tiru Kavala (@tirumala_rao_42902c822616).</description>
    <link>https://dev.to/tirumala_rao_42902c822616</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4244077%2F255ec4c3-3579-48f1-b42d-30418a9e8405.png</url>
      <title>DEV Community: Tiru Kavala</title>
      <link>https://dev.to/tirumala_rao_42902c822616</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tirumala_rao_42902c822616"/>
    <language>en</language>
    <item>
      <title>What a 13-agent Claude Code team taught me about agent harnesses</title>
      <dc:creator>Tiru Kavala</dc:creator>
      <pubDate>Mon, 16 Nov 2026 08:10:26 +0000</pubDate>
      <link>https://dev.to/tirumala_rao_42902c822616/what-a-13-agent-claude-code-team-taught-me-about-agent-harnesses-4j6l</link>
      <guid>https://dev.to/tirumala_rao_42902c822616/what-a-13-agent-claude-code-team-taught-me-about-agent-harnesses-4j6l</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://www.linkedin.com/pulse/what-13-agent-claude-code-team-taught-me-agent-harnesses-kavala-ws1vf/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;I run FinMoment, an AI bookkeeping pipeline I built, with a team of 13 Claude Code agents. The biggest lesson from operating it: the model is not the part that needs the most engineering. The harness around it is.&lt;/p&gt;

&lt;p&gt;By "harness" I mean everything that decides what an agent is allowed to do, how we know its work is finished, and where a human has to say yes. These are the rules I ended up with.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 1: "Done" is a check, not a feeling
&lt;/h2&gt;

&lt;p&gt;An agent will tell you a task is complete when it believes it is complete. That is not the same thing. Every task in my setup has a done-check: a concrete condition a machine can verify. The tests pass. The build succeeds. The change does what the task said it should.&lt;/p&gt;

&lt;p&gt;If the check fails, the task is not done, no matter how confident the summary sounds. This one rule removes most of the "it said it fixed it, but it didn't" frustration people have with coding agents.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 2: One job per agent, and a narrow set of tools
&lt;/h2&gt;

&lt;p&gt;Thirteen agents sounds like a lot, but each one is small. A narrow role means a short instruction set, fewer tools, and a smaller blast radius when it gets something wrong.&lt;/p&gt;

&lt;p&gt;It also makes failures easier to read. When something goes wrong, I know which agent produced it and which instructions to fix, instead of debugging one giant prompt that tries to do everything.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 3: The pipeline is the boss, not me
&lt;/h2&gt;

&lt;p&gt;I don't want to read every diff by hand. I want the pipeline to refuse bad work automatically. FinMoment runs on Next.js, Supabase and Vercel with GitHub Actions, and nothing ships unless the CI and deploy gates pass. Agents can propose changes all day; they cannot skip the gate.&lt;/p&gt;

&lt;p&gt;This comes straight from my years building large systems (about 15, including fraud-decisioning platforms). Zero-downtime deployments only worked there because the pipeline, not someone's mood on a Friday, decided what shipped. Agents need the same discipline, arguably more.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 4: Some decisions are human-only, by design
&lt;/h2&gt;

&lt;p&gt;There is a short list of actions no agent can take on its own: anything that approves financial data, touches money, or changes what customers see in production. These are hard gates, not polite suggestions in a prompt. The agent prepares everything, then stops and asks.&lt;/p&gt;

&lt;p&gt;This mirrors how FinMoment itself works. Uncertain fields go to a review queue, balances must reconcile, and a person approves before anything is exported. Nothing is guessed or auto-approved. The agents that build the product follow the same philosophy as the product.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 5: Make the state visible
&lt;/h2&gt;

&lt;p&gt;Agents lose context and sessions end. So the harness writes state where I can see it: what was done, what was checked, what failed, and what is waiting for a human. When I open my laptop I can see in a minute what happened and what needs me.&lt;/p&gt;

&lt;p&gt;I use the same skeleton for smaller agents I run every day: one turns chat payment screenshots into a ledger and day book, one reviews daily farm pond logs and raises alerts, one hunts for market problems and keeps a tracker. Different jobs, same pattern: a narrow role, a done-check, a log, and a human gate for anything that matters.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 6: Treat agent access to your tools like an API key
&lt;/h2&gt;

&lt;p&gt;FinMoment exposes an MCP endpoint so Claude can drive it. That is powerful, and it deserves the same care as any credential: scoped access, row-level security in Postgres, and human approval before anything important is committed. An agent connected to your tools is a new user of your system. Design for it like one.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this means if you're adding agents to your business
&lt;/h2&gt;

&lt;p&gt;You don't need 13 agents. You probably need one or two, with clear done-checks, a pipeline that blocks bad work, and a short list of decisions only a person can make. That is where most of the reliability comes from, far more than from picking the "best" model.&lt;/p&gt;

&lt;p&gt;That's the work I do: AI automation for finance and document workflows, Claude agents and agent harnesses with approval gates, and AWS/platform architecture. If you'd like to see a harness like this running on a real product, I'm happy to walk you through FinMoment on a call.&lt;/p&gt;

&lt;p&gt;You can find me on Upwork: &lt;a href="https://www.upwork.com/freelancers/%7E0146d40003db2cffd3" rel="noopener noreferrer"&gt;https://www.upwork.com/freelancers/~0146d40003db2cffd3&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>devops</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Zero-downtime deploys without Kubernetes: the ECS setup I keep coming back to</title>
      <dc:creator>Tiru Kavala</dc:creator>
      <pubDate>Mon, 16 Nov 2026 06:53:27 +0000</pubDate>
      <link>https://dev.to/tirumala_rao_42902c822616/zero-downtime-deploys-without-kubernetes-the-ecs-setup-i-keep-coming-back-to-2g29</link>
      <guid>https://dev.to/tirumala_rao_42902c822616/zero-downtime-deploys-without-kubernetes-the-ecs-setup-i-keep-coming-back-to-2g29</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://www.linkedin.com/pulse/zero-downtime-deploys-without-kubernetes-ecs-setup-i-keep-kavala-jc0lf/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Most small teams I talk to think zero-downtime deploys need Kubernetes. They don't. For a lot of products, Docker containers on AWS ECS behind a load balancer, plus a disciplined pipeline, gets you there with far less to operate.&lt;/p&gt;

&lt;p&gt;I spent about 15 years building large systems, including fraud-decisioning platforms where a bad deploy was not an option, and I've been hands-on with AWS since 2014. Here is the setup I keep coming back to, and the reasons behind each piece.&lt;/p&gt;

&lt;h2&gt;
  
  
  The goal: users never notice a release
&lt;/h2&gt;

&lt;p&gt;Zero-downtime is simple to state. While a new version goes live, every request still gets a correct answer. No maintenance page, no dropped sessions, no "try again in five minutes."&lt;/p&gt;

&lt;p&gt;Three things make that possible: something that only sends traffic to healthy copies of your app, a way to start the new version next to the old one, and a fast, boring way back if the new version is wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 1: The same Docker image, everywhere
&lt;/h2&gt;

&lt;p&gt;Build one image per commit and promote that exact image from test to production. Don't rebuild for each environment. If the image that passed your tests is not the image that runs in production, your tests didn't test production.&lt;/p&gt;

&lt;p&gt;Configuration that differs between environments comes in from outside (environment variables or a secrets store), never baked into the image.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 2: Health checks that mean something
&lt;/h2&gt;

&lt;p&gt;ECS and the load balancer decide which containers get traffic based on health checks. A check that only returns "200 OK" because the web server is up is close to useless. The app can be up and still unable to reach its database.&lt;/p&gt;

&lt;p&gt;A good health endpoint answers one question honestly: can this copy serve a real request right now? Keep it fast, check the dependencies that matter, and don't let it do expensive work. If the new version can't pass it, it never receives traffic, and that is exactly what you want.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 3: Rolling deploys with a safety margin
&lt;/h2&gt;

&lt;p&gt;ECS rolling updates start new tasks before stopping old ones. The two settings that matter are how many tasks must stay healthy during the deploy and how many extra tasks you allow on top. Set them so you never drop below the capacity you need, even for a minute.&lt;/p&gt;

&lt;p&gt;Then give new tasks time to warm up before the load balancer judges them, and let old tasks finish the requests they're already handling before they stop (connection draining). Skipping those two details is the most common reason "rolling" deploys still drop requests.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 4: Database changes that work with both versions
&lt;/h2&gt;

&lt;p&gt;During a rolling deploy, old and new code run at the same time against the same database. So every schema change has to work with both. Add a column first, deploy code that uses it, and only remove old columns in a later release. It feels slower. It is much faster than a midnight rollback that can't undo a dropped column.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 5: Rollback is a button, not a project
&lt;/h2&gt;

&lt;p&gt;Because every release is just an image tag, rolling back means pointing the service at the previous tag. ECS's deployment circuit breaker can do this automatically when new tasks keep failing their health checks. Practise it before you need it. A rollback path you have never used is a hope, not a plan.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 6: The pipeline decides what ships
&lt;/h2&gt;

&lt;p&gt;I've built this with Jenkins and with GitHub Actions; the tool matters less than the rule. Build, test, scan, deploy to a test environment, check it, then promote. Nobody deploys from a laptop. The same rule runs my own product today: FinMoment, an AI bookkeeping pipeline on Next.js, Supabase and Vercel, ships only when the CI and deploy gates pass, even though a team of Claude Code agents operates it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Piece 7: Watch the release, not just the server
&lt;/h2&gt;

&lt;p&gt;After a deploy, look at error rates and latency, not only CPU. I've used Splunk, Kibana and New Relic for this; what matters is having a dashboard or alert that tells you within minutes whether the new version is worse than the old one.&lt;/p&gt;

&lt;h2&gt;
  
  
  When you do need more
&lt;/h2&gt;

&lt;p&gt;This setup won't fit every case. If you run dozens of services with complex scheduling needs, a bigger platform may earn its keep. But many teams adopt that complexity long before they need it, and then spend their time operating the platform instead of the product.&lt;/p&gt;

&lt;p&gt;That's the kind of work I do: AWS and platform architecture, CI/CD pipelines with real gates, and AI automation for finance and document workflows, including Claude agents with human approval gates. If you want a second pair of eyes on your deploy setup, or want to see how I run FinMoment, I'm happy to walk you through it on a call.&lt;/p&gt;

&lt;p&gt;You can find me on Upwork: &lt;a href="https://www.upwork.com/freelancers/%7E0146d40003db2cffd3" rel="noopener noreferrer"&gt;https://www.upwork.com/freelancers/~0146d40003db2cffd3&lt;/a&gt;&lt;/p&gt;

</description>
      <category>aws</category>
      <category>devops</category>
      <category>docker</category>
      <category>cicd</category>
    </item>
    <item>
      <title>What a 13-agent Claude Code team taught me about agent harnesses</title>
      <dc:creator>Tiru Kavala</dc:creator>
      <pubDate>Mon, 16 Nov 2026 06:42:57 +0000</pubDate>
      <link>https://dev.to/tirumala_rao_42902c822616/what-a-13-agent-claude-code-team-taught-me-about-agent-harnesses-1fp1</link>
      <guid>https://dev.to/tirumala_rao_42902c822616/what-a-13-agent-claude-code-team-taught-me-about-agent-harnesses-1fp1</guid>
      <description>&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://www.linkedin.com/pulse/what-13-agent-claude-code-team-taught-me-agent-harnesses-kavala-ws1vf/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;I run FinMoment, an AI bookkeeping pipeline I built, with a team of 13 Claude Code agents. The biggest lesson from operating it: the model is not the part that needs the most engineering. The harness around it is.&lt;/p&gt;

&lt;p&gt;By "harness" I mean everything that decides what an agent is allowed to do, how we know its work is finished, and where a human has to say yes. These are the rules I ended up with.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 1: "Done" is a check, not a feeling
&lt;/h2&gt;

&lt;p&gt;An agent will tell you a task is complete when it believes it is complete. That is not the same thing. Every task in my setup has a done-check: a concrete condition a machine can verify. The tests pass. The build succeeds. The change does what the task said it should.&lt;/p&gt;

&lt;p&gt;If the check fails, the task is not done, no matter how confident the summary sounds. This one rule removes most of the "it said it fixed it, but it didn't" frustration people have with coding agents.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 2: One job per agent, and a narrow set of tools
&lt;/h2&gt;

&lt;p&gt;Thirteen agents sounds like a lot, but each one is small. A narrow role means a short instruction set, fewer tools, and a smaller blast radius when it gets something wrong.&lt;/p&gt;

&lt;p&gt;It also makes failures easier to read. When something goes wrong, I know which agent produced it and which instructions to fix, instead of debugging one giant prompt that tries to do everything.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 3: The pipeline is the boss, not me
&lt;/h2&gt;

&lt;p&gt;I don't want to read every diff by hand. I want the pipeline to refuse bad work automatically. FinMoment runs on Next.js, Supabase and Vercel with GitHub Actions, and nothing ships unless the CI and deploy gates pass. Agents can propose changes all day; they cannot skip the gate.&lt;/p&gt;

&lt;p&gt;This comes straight from my years building large systems (about 15, including fraud-decisioning platforms). Zero-downtime deployments only worked there because the pipeline, not someone's mood on a Friday, decided what shipped. Agents need the same discipline, arguably more.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 4: Some decisions are human-only, by design
&lt;/h2&gt;

&lt;p&gt;There is a short list of actions no agent can take on its own: anything that approves financial data, touches money, or changes what customers see in production. These are hard gates, not polite suggestions in a prompt. The agent prepares everything, then stops and asks.&lt;/p&gt;

&lt;p&gt;This mirrors how FinMoment itself works. Uncertain fields go to a review queue, balances must reconcile, and a person approves before anything is exported. Nothing is guessed or auto-approved. The agents that build the product follow the same philosophy as the product.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 5: Make the state visible
&lt;/h2&gt;

&lt;p&gt;Agents lose context and sessions end. So the harness writes state where I can see it: what was done, what was checked, what failed, and what is waiting for a human. When I open my laptop I can see in a minute what happened and what needs me.&lt;/p&gt;

&lt;p&gt;I use the same skeleton for smaller agents I run every day: one turns chat payment screenshots into a ledger and day book, one reviews daily farm pond logs and raises alerts, one hunts for market problems and keeps a tracker. Different jobs, same pattern: a narrow role, a done-check, a log, and a human gate for anything that matters.&lt;/p&gt;

&lt;h2&gt;
  
  
  Rule 6: Treat agent access to your tools like an API key
&lt;/h2&gt;

&lt;p&gt;FinMoment exposes an MCP endpoint so Claude can drive it. That is powerful, and it deserves the same care as any credential: scoped access, row-level security in Postgres, and human approval before anything important is committed. An agent connected to your tools is a new user of your system. Design for it like one.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this means if you're adding agents to your business
&lt;/h2&gt;

&lt;p&gt;You don't need 13 agents. You probably need one or two, with clear done-checks, a pipeline that blocks bad work, and a short list of decisions only a person can make. That is where most of the reliability comes from, far more than from picking the "best" model.&lt;/p&gt;

&lt;p&gt;That's the work I do: AI automation for finance and document workflows, Claude agents and agent harnesses with approval gates, and AWS/platform architecture. If you'd like to see a harness like this running on a real product, I'm happy to walk you through FinMoment on a call.&lt;/p&gt;

&lt;p&gt;You can find me on Upwork: &lt;a href="https://www.upwork.com/freelancers/%7E0146d40003db2cffd3" rel="noopener noreferrer"&gt;https://www.upwork.com/freelancers/~0146d40003db2cffd3&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>devops</category>
      <category>productivity</category>
    </item>
    <item>
      <title>AI shouldn't guess with money: 5 design rules from building an AI bookkeeping pipeline</title>
      <dc:creator>Tiru Kavala</dc:creator>
      <pubDate>Sat, 14 Nov 2026 07:40:21 +0000</pubDate>
      <link>https://dev.to/tirumala_rao_42902c822616/ai-shouldnt-guess-with-money-5-design-rules-from-building-an-ai-bookkeeping-pipeline-3d0h</link>
      <guid>https://dev.to/tirumala_rao_42902c822616/ai-shouldnt-guess-with-money-5-design-rules-from-building-an-ai-bookkeeping-pipeline-3d0h</guid>
      <description>&lt;p&gt;Most AI demos for finance look great until the numbers have to add up. A model that is 97% right on bank statements is still wrong on three lines out of a hundred, and in bookkeeping one wrong line means the books don't close.&lt;/p&gt;

&lt;p&gt;I spent about 15 years building large systems, most recently as a solution architect and engineering manager on fraud-decisioning platforms at Capital One. When I started building FinMoment, an AI bookkeeping pipeline, I brought one rule from that world with me: the system is allowed to be unsure, but it is never allowed to guess.&lt;/p&gt;

&lt;p&gt;Here are the five design rules that came out of it. They apply to any AI agent that touches money, invoices or documents people rely on.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Let the model read, but let code decide
&lt;/h2&gt;

&lt;p&gt;The pipeline is: upload, then OCR, then extraction, then a review queue, then deterministic validation, then human approval, then export (CSV, Excel or Tally XML).&lt;/p&gt;

&lt;p&gt;The language model does the parts it is good at: reading messy statements and receipts, and turning them into structured fields. It does not decide whether the result is correct. That is done by plain code with rules that can't be talked out of an answer.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Make the numbers prove themselves
&lt;/h2&gt;

&lt;p&gt;The most useful check is also the simplest: balances must reconcile. Opening balance plus credits minus debits has to equal the closing balance, for every page of every statement. If it doesn't, something was misread, skipped or duplicated, and the run stops there instead of producing a nice-looking file that's wrong.&lt;/p&gt;

&lt;p&gt;A check like this catches errors a confidence score never will, because the model can be very confident and still be wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Send uncertainty to a human, field by field
&lt;/h2&gt;

&lt;p&gt;When a field is unclear (a smudged amount, an ambiguous date, a merchant name the model can't place), it goes to a review queue instead of being filled in with a best guess. The person sees exactly which fields need attention, not a whole document to re-check.&lt;/p&gt;

&lt;p&gt;This is the difference between "AI that saves time" and "AI that creates a second job of checking the AI".&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Link every number back to its source
&lt;/h2&gt;

&lt;p&gt;Every transaction is labelled and linked to the line it came from. When an accountant asks "where did this come from?", the answer is one click away. Without that, nobody can audit the output, and output nobody can audit is output nobody can trust.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Agents can drive the system, but humans approve
&lt;/h2&gt;

&lt;p&gt;FinMoment has an MCP endpoint, so Claude can drive the pipeline. The final approval step is human-only. I run the product with a team of 13 Claude Code agents, and the same pattern applies there: done-checks, CI and deploy gates, and approval gates that only a human can pass.&lt;/p&gt;

&lt;p&gt;Agents are fast and tireless. They are also confidently wrong sometimes. Gates are how you keep the speed without inheriting the mistakes.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this means if you're adding AI to a finance or document workflow
&lt;/h2&gt;

&lt;p&gt;Before picking a model, write down three things:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;What check proves the output is right? (Totals, reconciliations, cross-field rules.)&lt;/li&gt;
&lt;li&gt;Where does uncertain data go? (A queue with the exact fields, not a full re-review.)&lt;/li&gt;
&lt;li&gt;Who approves, and what do they see? (The source, the extracted value and the reason it was flagged.)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you can answer those, the model choice matters much less than people think.&lt;/p&gt;

&lt;p&gt;I help teams build this kind of AI automation: document and invoice extraction, bookkeeping workflows, and Claude agents with approval gates and audit trails, usually on AWS or a Next.js/Supabase stack. If you're working on something similar, I'm happy to walk you through FinMoment on a call. You can find me on Upwork: &lt;a href="https://www.upwork.com/freelancers/%7E0146d40003db2cffd3" rel="noopener noreferrer"&gt;https://www.upwork.com/freelancers/~0146d40003db2cffd3&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally published on &lt;a href="https://www.linkedin.com/pulse/ai-shouldnt-guess-money-5-design-rules-from-building-pipeline-kavala-5iodf/" rel="noopener noreferrer"&gt;LinkedIn&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>automation</category>
      <category>architecture</category>
    </item>
  </channel>
</rss>
