Memory gives agents context. Observability shows traces. Orchestration moves tasks forward. But none of that answers the question every engineering manager asks when an AI agent screws up: what did we actually expect to happen, and why didn't it? AccInt, a new project emerging from Hacker News on June 13, is trying to close that gap with something its creators call a Work Model—a persistent ledger that records commitments, verifies them against reality, and compounds learned judgment over time.

The Core Insight: Settlement, Not Vibes

The project's homepage cuts through the noise with surgical precision. Most agent frameworks focus on making AI do more stuff. AccInt focuses on what happens after the doing—specifically, whether the commitment matched the outcome. Every action gets recorded as a written commitment before it executes, then verified against what actually came back. The system tracks which paths deserve reuse and scores them by results. It's accountability for autonomous agents, built from the ground up to run locally on hardware you control. No cloud brain to rent. No API key to leak.

How It Actually Works

The technical architecture leans hard into local-first principles: a pure Rust binary paired with a SQLite file holds everything. The AI engine itself is swappable—AccInt's creators claim the same Work Model runs under two different engines today, which matters for teams worried about vendor lock-in. Under the hood, it borrows from ColBERT late-interaction retrieval, JEPA-style state transitions, and what they describe as pedagogical reinforcement learning to update the scored model underneath. The installer probes your VRAM, RAM, and disk space, then picks an appropriate rung on their hardware ladder: Qwen3 8B AWQ for beefy NVIDIA cards, scaling down through smaller quantizations until it hits a text-only floor for underpowered setups.

Real Numbers From Live Operation

The AccInt operator behind its own landing page provides live telemetry as proof of concept. As of June 12, 2026: 2,778 entities held in the local Work Model with 537 marked as runnable, 20,667 events recorded on the append-only ledger, and 1,931 outcomes scored against reality—372 self-graded but held at weak priors to avoid overconfidence. Their last 50 runs hit 94% success, though they stress that belief never counts as reality in their system. The built-in eval harness claims recall@5 = 1.000 against exact brute-force across the live model.

First Wedge: Coding Agents Already Doing Real Work

The product isn't aimed at teams evaluating whether to use AI agents. It's for teams already handing work to Cursor, Codex, Claude, OpenCode, or MCP-connected tools and asking harder questions about accountability. AccInt positions its first wedge as software factories, technical agencies, research loops, sales pipelines, and finance work where a bad action has a cost. The demo shows an agency recruiting workflow: sourcing candidates, drafting personalized outreach held for human approval before sending, then tracking outcomes to credit what worked—while refusing actions like bulk blast emails that violate consent floors they set programmatically.

Early Access Philosophy

Invites go out in order, with the team explicitly prioritizing strongest fit first. They're not chasing volume; they're after teams already putting agents near real work who understand the difference between a prettier prompt and accountable learning. The whitepaper contains the full technical argument and math for those who want to dig into the JEPA transitions and recursive language model routing policy.

Key Takeaways

  • AccInt builds an append-only ledger of agent commitments, settling each against verified outcomes rather than trusting confidence scores alone
  • Runs entirely local: one Rust binary, one SQLite file, no cloud control plane or API keys floating around
  • The Work Model is engine-agnostic—swap the AI underneath without losing what's been learned
  • Live telemetry shows 94% success on recent runs with over 20,000 events recorded since deployment

The Bottom Line

This is the kind of project that makes you wonder why no one built it sooner. Memory and observability are table stakes now; what separates useful agent work from dangerous agent work is knowing whether your system learned anything or just got luckier this time. AccInt's ledger-first approach won't be for everyone—it's opinionated, requires operational maturity, and the early access queue means you'll wait—but for teams already deep in AI-assisted development and hungry for accountability rather than more capability, this might be exactly the substrate they've been missing.