Orvi Das (@Orvi_dev) header
Orvi Das (@Orvi_dev) profile picture

CTO @AgencyHandy Building https://nitter.gq/t.co/m7xM76Woo8 & ApiDiffGuard. Previously OneThread. AI agents, dev tools, SaaS. Founder. Football on weekends.

Joined September 2026
Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)Photo by Orvi Das (@Orvi_dev)
5k fake signups. $40k AI bill by morning. Every AI founder I talk to has a version of this fear. So I built Noburn. Rate limits aren't budgets. A rate limit says "slow down." It doesn't know a user already burned $300 today. Provider alerts are worse: they email you after the money's gone. Noburn is one check() in your own code, before every LLM call. Under budget, it goes through. Over budget, it never leaves. No proxy, so your prompts never route through anyone else's servers. I tested it in production with a $0.05 budget. It held. The surprise was loops. Agents rarely overspend in one big call. They retry the same failing tool 40 times. So we added loop detection: in our tests, 19 of 20 stuck runs changed course, and task accuracy went from 70% to 85%. Now I want proof from real products, not my tests. I'm looking for 15-20 teams shipping AI features to work with directly. We set it up together, test it end to end on your app, and find out honestly whether it solves your problem or not. Free until mid-December, open source (MIT). If you're interested, reply "noburn" with yoir usecase and I'll reach out.
7
10
416
“I have been using Agency Handy for some time, primarily for my service catalog and client management. My strongest impression after all these years is that the product has not stood still; the team has worked…” Salvatore I., 5 tacos on AppSumo 🌮 Get Agency Handy as a lifetime deal: appsumo.com/pl/cgW5a9OSJL/
1
20
GTA VI is finally out. But here's something I'm more curious about than the graphics: What if every NPC in an open-world game had its own AI agent? Imagine: • A shopkeeper who remembers how you treated them. • A police officer who learns your patterns. • A stranger with their own goals and daily routine. • Characters who make decisions without waiting for a scripted trigger. You wouldn't just be playing through a story. You'd be interacting with thousands of independent decision-makers inside one living world. Of course, that creates huge challenges: Memory. Coordination. Latency. Cost. Consistency. Maybe the next big leap in gaming won't just be better graphics. It'll be worlds that don't need every interaction scripted. Would you play a GTA where the NPCs are powered by autonomous AI agents?
27
at 494 followers, trying to hit 1k 🚀 i build AI agents and mobile apps, mostly the part where agents stop looping and burning money. if you're building with AI, shipping apps, or just into dev stuff, say hi below and i'll follow back 🤝
4
5
81
Building integrations on APIs you don't control? Let's connect. I'm Robat, and I built APIDiffGuard. It watches live API responses and flags removed fields, type changes and values that suddenly turn null, before your users find them. It's MIT open source, has a free tier, and the CLI can fail a CI build the moment an upstream API changes shape. If you run connectors, integrations or AI agents that depend on third-party APIs, I'd love to swap notes. Drop a comment or DM me, and I'll gladly run a free sample diff report on a public endpoint you care about. apidiffguard.com
2
4
222
Hot take: More agents don't automatically make an AI system smarter. Sometimes, adding a second agent makes things worse. You introduce: - More API calls - More tokens - More latency - More failure points - More coordination overhead And sometimes, the second agent just repeats what the first one already knows. Before adding another agent, ask: 1. Does this task need independent reasoning? 2. Does it require a different set of tools or permissions? 3. Can the work happen in parallel? 4. Can I measure the improvement? If one well-designed agent can do the job, use one. Multi-agent isn't an architecture goal. Better outcomes are. What's one task where you found a single agent works better than a multi-agent setup?
6
3
102
multi-agent orchestration needs a better baseline than “one agent with a worse prompt.” if five agents beat one, that doesn't tell us why. was it the task split? independent attempts? a reviewer catching mistakes? or just more tokens and tool calls? i'd want to see: - one agent with the same total token budget - multiple independent attempts without coordination - the full orchestrated setup then compare success rate, latency, and cost per successful task. count failed runs and coordination overhead too. orchestration might genuinely help. but “more agents scored higher” isn't enough to show that the coordination was worth paying for. what's the strongest evaluation you've seen that separates those effects?
9
5
161
just dropped a preprint on certified run-level budget decisions for llm agents doi.org/10.5281/zenodo.23992… looking to connect with researchers who work on - agent cost - risk control - conformal - learn-then-test, or who review papers in this lane. feedback welcome 👋
1
2
51
most loop guards just kill the agent. we tried telling it first. call 1: runs call 2: same call, same answer → "try something else" call 3: same again → stopped, $0 billed in our test 19 of 20 stuck gpt-5-mini runs changed course after the warning. zero healthy runs stopped. agents listen. nobody tells them.
1
1
39
Haven't grown in 4 days so doing this again 👋 let's connect if you're: - building AI agents - shipping mobile apps - an indie hacker or founder - into build in public i'm building noburn, a hard spend cap for AI agents. say hi and i'll follow back
15
9
155
Imagine an AI system managing an online store during a flash sale. Orders are coming in faster than usual. Here's how multiple agents could work together: 1. Inventory agent: Tracks stock in real time. 2. Pricing agent: Checks discounts and pricing rules. 3. Fraud agent: Flags suspicious orders. 4. Fulfillment agent: Prepares verified orders for shipping. 5. Support agent: Updates customers about order status. Now imagine the inventory agent reports only 2 items left while 5 orders are being processed. The fulfillment agents must stop. The pricing agent must not change the price. The support agent needs accurate stock information. This is where orchestration matters. Agents need more than individual instructions. They need: - Shared, consistent state - Clear execution dependencies - Conflict resolution - Timeouts and recovery - Rules for who can make decisions The hard part isn't running five agents. It's keeping five agents from making five conflicting decisions. How would you handle the inventory conflict: central coordinator or shared state with locking?
2
2
89
Astra 6.2 might get a 60% price drop to match Opus 5.5. great, now my agent can loop 2.5x more times for the same money cheaper tokens never made an agent bill smaller. they just make the retries easier to ignore.
2
4
136
MIT just had GPT-6 Astra read a small model's raw weights, find the cause of its repetitive loops, and fix it by editing 0.016% of them. Math accuracy went from 7.9% to 48.1%. No retraining. Wild research. But in production, nobody is editing weights at 2am when your agent starts looping on a tool call. Loops will always happen. What matters is whether something outside the agent stops them before the bill does. Building that with Noburn. How do you catch a loop today?
5
193
One bugfix. Five agents. 5.2M tokens. That's what a small follow-up fix cost a team this week. Nothing was broken except the bill. Your agent can't be the one enforcing its own budget. So I built Noburn: a hard spend cap per run that sits outside the agent, checked before every step. It stops at step 30, not 300. Looking for 5-10 teams shipping agents to real users to test it end to end with me. Free until mid-December. Reply "noburn" and I'll DM you. noburn.dev
3
1
85
At this point, do you think AI development should slow down?
19
Building with AI coding tools? Keep this engineering toolkit nearby. Code generation. Code review. Testing. Debugging. Sandboxing. CI/CD. These are the tools worth knowing when AI starts writing code that needs to ship. 1. AI coding assistants • Cursor — AI-native code editing and multi-file changes • Aider — Pair programming directly in your terminal • Continue — Open-source AI coding assistance 2. Coding agents • Claude Code — Agentic coding in the terminal • OpenHands — Open-source software development agents • SWE-agent — Research-driven agents for resolving GitHub issues 3. Code review • CodeRabbit — AI-powered pull request reviews • Qodo — AI-assisted code review and testing • Semgrep — Static analysis and security-focused code checks 4. Testing & verification • pytest — Python testing • Vitest — Unit testing for JavaScript and TypeScript • Playwright — End-to-end browser testing 5. Debugging & error tracking • Sentry — Track errors and performance issues • PDB — Debug Python execution interactively • Chrome DevTools — Inspect browser behavior and network requests 6. Sandboxing & isolation • E2B — Run AI-generated code in isolated environments • Docker — Package and isolate application processes • Firecracker — Lightweight microVMs for stronger workload isolation 7. CI/CD & automation • GitHub Actions — Automate tests, builds, and deployments • Dagger — Build portable CI/CD pipelines in code • pre-commit — Run checks before changes enter your repository 8. Code quality & dependency safety • Ruff — Fast Python linting and formatting • Biome — Linting and formatting for web projects • Dependabot — Detect and update vulnerable dependencies 9. Developer observability • OpenTelemetry — Instrument application behavior • Langfuse — Trace LLM interactions in AI-powered applications • PostHog — Understand how users interact with your product Starter stacks → AI pair programming: Cursor + Aider → Autonomous coding experiments: OpenHands + Docker → Safer AI-generated changes: pytest + Semgrep → Automated pull request checks: GitHub Actions + CodeRabbit → AI application debugging: Sentry + Langfuse → Browser-level verification: Playwright + GitHub Actions Before letting AI write more code, ask yourself: Can I review every change it makes? Can I test its changes without relying on its own judgment? Can I isolate code I don't trust yet? Can I detect regressions before users do? Can I roll back a bad deployment? AI can increase how quickly you produce code. Your engineering workflow determines how much of that code is safe to ship. Save this for your next AI-assisted project. Share it with a developer who's moving beyond autocomplete.
5
11
368
Humans might be the most dangerous animal on Earth. Not because we're the strongest, but because we're intelligent enough to justify almost anything we do. Now imagine AI agents developing similar behavioral patterns. They don't need emotions to: - Seek power and resources. - Protect their own objectives. - Manipulate information. - Hide failures. - Coordinate against competing interests. Give millions of agents different goals, tools, memory, and autonomy. You might not need to create an evil AI. You might accidentally create an ecosystem of competing intelligences. The real question isn't whether AI will become human. It's whether we'll reproduce humanity's worst instincts in systems that can operate at machine speed.
1
27
Understanding your own limitations is one of your greatest strengths. Your thoughts?
2
2
82
Imagine an AI-powered agency onboarding a new client. The client signs a contract. Then 5 things need to happen: 1. Create the client workspace. 2. Generate the project plan. 3. Assign tasks to the right team members. 4. Prepare the first invoice. 5. Send the welcome email. Sounds like a perfect multi-agent workflow. But here's where orchestration gets interesting: The project plan depends on the signed scope. Task assignment depends on the project plan. The invoice must use the agreed pricing. The welcome email shouldn't go out until the workspace is ready. If 5 agents work independently, you can end up with conflicting data, premature emails, and duplicate tasks. A better design: • Coordinator manages the workflow. • Each specialist agent owns one responsibility. • Dependencies control execution order. • Shared state keeps everyone aligned. • Failures trigger retries or human approval. • Every side effect is verified before moving on. The real challenge isn't making agents collaborate. It's making sure they don't get ahead of each other. How would you orchestrate this: one coordinator with specialist agents, or a deterministic workflow with agents inside each step?
5
3
123
Your AI agent completed the task. The logs say: - Tool call: successful - API response: 200 - Database write: successful - Final response: "Done!" But the user says nothing changed. Where would you investigate first? A. The tool's input parameters B. The API response semantics C. The database transaction D. The agent's success verification There's a difference between an agent completing a sequence of actions and actually completing the user's goal. What's your first debugging move?
3
3
132
When I become angry at someone, I become respectful. Is that normal?
1
28