Dr. Tali Režun @talirezun

Non-developer, 30 years of building companies, shipped all of this in the last 6 months using AI agents: Two open-source frameworks: The Curator, a local second brain with MCP server that compounds knowledge across agent sessions, github.com/talirezun/the-… ØØT, an open-source org design framework for AI-augmented companies, github.com/talirezun/oot-… Two production SaaS products live: @luminawidget (AI chat widget platform, built for under $100 in credits) and Moj AI (multi-agent RAG for legal navigation) One MCP server for the tourism industry, one AI avatar platform in active client development, 7,000 lines of code, 300+ agent sessions Never written a line of code except HTML and CSS. All orchestrated. The bottleneck was never the code, it was context and vision.

@neogoose_btw Full breakdown of all 12 components: talirezun.substack.com/p/blueprint-of…

This is exactly the right test to run. "Native harness is better" was always an assumption dressed as a principle. The real variable isn't who built the harness, it's what the harness is actually doing with the model: how it manages context, how it picks tools, how it handles errors before they compound, whether it verifies output or just declares success. Those decisions explain most of the performance and cost difference, not native vs alternative. Same model, completely different outcome depending on those choices. Spent 18 months mapping this across real builds. Ended up with 12 specific components that account for most of the variation. The tools finding is the most counterintuitive one: fewer tools consistently beats more tools, regardless of harness. Vercel's v0 team removed 80% of available tools and got better results.

Probably not a dev in a legacy meaning. But you can call yourself a founder, a product architect, an orchestrator. I've shipped many production SaaS products, MCPs, two open-source frameworks without writing a line of code. The question isn't whether you can claim the dev title. It's whether the product works and solves a real problem. Nobody asks the CEO of a construction company if he poured the concrete himself.

@emollick github.com/talirezun/oot-…

The "comfortable" framing is what worries me actually. MIT NANDA clocked 95% of enterprise GenAI pilots at zero measurable P&L impact. And the METR study is the real gut punch: experienced developers were 19% slower on real tasks using AI tools while self-reporting they were 20% faster. That's a 39 point perception gap. The problem isn't workflow integration. It's that companies think they're adopting while the numbers say otherwise. The firms that get past the waypoint aren't the ones with better tools. They're the ones that took the resistance layer seriously first, as a design problem, not a training problem. Built an open-source framework specifically around this, including a mandatory governance gate before any "AI replaces human" decision executes. The Klarna case is encoded into it as a blocking test, not a case study.

@RoundtableSpace github.com/talirezun/the-…

Good execution stack. But all 10 of these share one structural flaw, they're stateless. Every session starts from zero. No memory of what was decided, what failed, what patterns worked. The one repo I'd add to this list is one that fixes that: The Curator. Local knowledge graph with an MCP server, agents can read and write to it directly during sessions. Six months in it's the most valuable thing in my stack. Because the automations above are only as good as the context you can feed them. Knowledge compounds. Tasks don't.

Just did exactly this. M3 Max, 32GB, dropped it into @lmstudio and ran first tests on Gemma 4 12B Coder. Honestly surprised by how well it handles focused coding tasks. Not Fable-level on complex reasoning but for contained, well-specified problems the gap is smaller than I expected. Will report back after more serious sessions but the first impression is solid. The "keep it in storage" advice is good even if you don't plan to use it today. You might need it on a Friday afternoon.

The incentive trap is correctly identified — and we already saw it compress from theory to practice in four days. Fable's shutdown is the proof of concept one level below AGI: a single directive and every non-American workflow lost frontier access overnight. The firms that had it didn't get better. The firms without it got measurably worse. That's not competition — that's manufactured advantage. The economic argument for access isn't just fairness: distributed frontier intelligence creates new markets, new companies, categories that didn't exist before. Captured frontier intelligence extracts from existing markets faster. One model expands the pie. The other consolidates who gets to eat it. The open question is whether any incentive structure can make labs prefer expansion over capture — or whether we're now past the point where that choice is theirs to make.

@MatthewBerman Wrote about this transition from single-agent to orchestrated multi-agent builds here — the "symphony vs solo" framing covers exactly why this breaks and how to fix it: open.substack.com/pub/talirezun/…

The simultaneous deploy conflict isn't a GitHub problem. It's an orchestration architecture problem. If multiple agents can independently push, you've given them each a baton — but there's no conductor. The fix: worker agents never touch the shared repo directly. They operate in isolated environments, complete their tasks, and hand off to a single orchestrator. The orchestrator owns the deploy queue, sequences the merges, and is the only thing that ever talks to main. Running 177 parallel tasks this way — zero conflicting pushes. The workers never knew main existed. What does your current handoff look like between task completion and the push step?

@garrytan github.com/talirezun/the-…

GBrain as a retrieval layer makes sense — just-in-time context from a large corpus is a real problem worth solving at the infrastructure level. The architecture I've been exploring with The Curator is a different layer: instead of chunking and retrieving from existing documents, agents read sources and write a structured wiki from scratch — entities, concepts, summaries, all interlinked and updated rather than duplicated with each new source. The distinction that matters operationally: retrieval gives agents what they need now. Knowledge compounding builds what they'll need six months from now. Both layers seem worth having in a mature stack. Curious whether GBrain handles continuous knowledge ingestion or is optimized primarily for static corpora?

Full blueprint 👇 X Article (condensed): x.com/talirezun/stat… Full Substack piece: open.substack.com/pub/talirezun/…

Dr. Tali Režun @talirezun

6 hours ago

x.com/i/article/2066…

1 0 0 42 0

x.com/i/article/2066…

Vercel's v0 team removed 80% of available tools from their coding agent. Performance improved. This is what most orchestrators miss about how frontier coding agents actually work. A production agent harness has 12 components. The ones that trip people up: → Context Management: performance drops 30%+ when key content hits mid-window positions. The agent isn't dumb — it's losing the thread. → Error Handling: 10 steps × 99% success = 90% end-to-end. Errors compound invisibly. → Verification Loops: demo agents declare success. Production agents run the tests. → Tools: the minimum set for this step. Not everything the agent might ever need. I built the primitive version of this architecture — manually — before it had a name. 18 months of handoff files, MCP server stacks, and painful context ceiling experiences taught me the architecture from the inside. Full blueprint of all 12 components in the comments.

Genuinely useful resource — RAG is a fundamental building block. One practitioner note worth adding for anyone starting here: after two years of RAG experimentation I could not get below 20% hallucination rate in best-case scenarios on complex document retrieval. The core problem is the chunking step — when you split documents into chunks, you lose the surrounding context that gives meaning to the retrieved fragment. The model gets the right piece of text but not enough of what surrounds it to reason accurately. For production use cases where hallucination cost is high, I've moved to a simpler architecture: a library system that retrieves full documents based on the prompt and passes them whole into a large context window. No chunking, no retrieval pipeline, no hallucination from missing context. Works in two production apps. RAG has its place — especially for very large corpora where full-document retrieval is impractical. But for most small-to-medium knowledge bases, the large context window approach is simpler, more reliable, and easier to debug.

Alex captures the feeling correctly even if the scale is uncertain. But here's what history tells us: once humans experience a step-change in capability, they don't accept losing it. Not permanently. We found ways around every previous restriction — encryption, open-source software, distributed networks. The appetite doesn't disappear when access is revoked. It accelerates. The immediate effect of Friday's decision is not European surrender. It's a starting gun for every non-US AI lab, every open-weight model team, every decentralized compute project on earth. China will not slow down. Europe will not accept permanent dependency. The race just got more intense, not less. Fable may be restricted. The hunger it created is not.

Sundown at the sea. After a day like today — Fable down, governments pulling switches, everyone building fallback architectures at speed — this is what I needed. Some things in this world still move at their own pace. The sea doesn't care about export control directives. Breathe. Some things will sort themselves out.