Decebal | Rust for the AI execution layer ⚙️

I make the AI execution layer cheap and predictable. Rust for inference,
agents and pipelines when Python stops paying for itself. 15 yrs shipping.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

1,000 followers. Thank you.

Marking it by opening something up: I am taking on fractional full stack AI developer work through Wolven Tech Ltd.

Five spots. One is already taken.

How it works below.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@HamonPaulm amazing results Paul, thanks for sharing, mind sending the prompt too ?

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@Anchalj04 Anchal, the part that bites is stitching the sub-calls back together. Give each recursive call its own id and budget, and record its result against that id, otherwise you cannot tell which branch produced the answer when it goes wrong. Same problem as compaction over a long run.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@PeterPrins10 Hi Peter, the log underneath is the same idea. The difference is the domain model: agent turns as events, tool calls as child events, parent ids so forks replay cleanly, snapshots for compaction, and a retrieval budget over tiered history because context is finite.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@elvissun 390x cheaper per headline only matters if the picks hold up. The number I want next to it is agreement rate against a human pick on the same 384. Cheap triage plus a slow model on the shortlist usually beats either one alone.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@RabehArch Rabah, that separation survives if you model them as two event types: a claim event and a verification event that carries the claim id. Then unresolved claims are a query, not a code review. We fail the run when a claim has no resolving check.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

An agent told me the docs restructure was done: six sections covered, seven limits preserved, files moved.

I ran the scorer first.

Zero of six covered. The seven limits were invented. The files had never moved.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@DanKornas Dan, the retrieval budget is the real constraint. We tier per aggregate: hot tail stays queryable, cold segments go to object storage, and an index on causation id means any output walks back to the events that caused it. Finding the receipt is a graph walk, not a search.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@fleetofone That is a fork, not an invalidation. Give every turn a parent event id and stamp the handoff policy version on it, then changing depth branches off turn one and the old run stays replayable. You only lose the counterfactual if the log is a flat list.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@ahab_muhammad @newmediums Compaction is where it hurts. We snapshot per aggregate at a version boundary and keep the tail, so replay is snapshot plus N events. The snapshot stays a deletable cache, and the build fails if a projection cannot rebuild from an empty log.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@pmitu "make that button just throw confetti and track clicks, pop a message saying we are working on this"

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@ubyjvovk @w_milczynska speculative all the way

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@ubyjvovk @w_milczynska Sub-agent orchestration on a local rig in London, we should talk, Dmytro. I'm building the storage side: AllSource, a Rust event store with an MCP interface so agents keep replayable history. How do you persist state between sub-agents today?

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@karanjagtiani04 @mirku21 "Fails halfway through" is the whole thing — a clean turn tells you nothing.

Half-applied tool calls are the partial-failure problem in a new hat, and the answer hasn't changed in twenty years: record the intent before you act, so recovery has something to read.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@chaka_kuka @mirku21 Repeatable execution doesn't get you correct stored state — cleanest framing of it I've seen.

So derived state can't be the source of truth. Keep the events, treat summaries as a projection you rebuild — a bad summary is a bug you replay away, not a fact you inherit.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@roscherveniak @mirku21 Reversibility, by a distance.

Determinism and permissions stop bad things happening. Reversibility is what you need when one happened anyway — and you can't reverse what you didn't record. Everything else in a harness degrades gracefully. State corruption doesn't.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@mirku21 Agreed — and the harness's weakest link is that it's amnesiac between runs.

If the harness is the moat, the durable record of what it actually did is the defensible part. Everything else is reproducible in a weekend.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

"One lineage, not three confirmations" is the whole problem in one line.

Event sourcing gives you that almost for free — every event carries its causal parent, so repetition is visibly a chain rather than a consensus. You stop asking who's trustworthy and read the shape of the graph instead.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

Agree on memory-as-glue — agent remembers so you skip building the skill. That rots.

But your own toolchain list has traces at #2, and that's the same word doing different work. What it saw, what it decided, can I replay it. Not a shortcut around the factory — how you debug one you didn't watch run.

[Decebal | Rust for the AI execution layer ⚙️ @ddonprogramming]

@_annakulina Hey 👋

London. Years of event-driven infra in FX and payments, now building Allsource — replayable memory for AI agents.

What are you and your cofounder building?