AI Agents Claiming 90% Autonomy May Deliver Just 61.6% Usable Output

·3 views

A new framework called Proof-Adjusted Autonomy (PAA) argues that standard AI agent autonomy metrics are misleading because they only measure task completion, not whether results are evidenced, independently validated, and delivered on time. When four production-critical factors — autonomous execution, evidence coverage, validation pass rate, and on-time delivery — are multiplied together, a 90% autonomous agent realistically yields only 61.6% truly usable output. The gap is attributed to what Grant Thornton has called the 'AI Proof Gap', where organizations deploy AI faster than they can establish accountability for its outputs. Researcher Jason Wei's 'Verifier's Rule' further explains that AI capabilities naturally develop first in easily verifiable domains, leaving harder-to-verify production tasks exposed. The PAA framework contends that AI does not eliminate the cost of work but shifts it from producing outputs to proving those outputs are correct.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer Audits AI Slack Workflow, Finds One Safety Check Never Triggered

A developer granted an AI system direct permission to post messages to their team's Slack workspace and built three verification checks into the workflow to ensure accountability. Upon auditing session logs on July 9, 2026, they discovered that one of the three checks had never been triggered even once. A second check also behaved differently from how the developer had originally described it, revealing a gap between assuming a system works and being able to verify what it actually did. The AI accessed Slack via MCP tools, and one safeguard required it to resolve channel names through a search function rather than using raw, human-unreadable channel IDs. The audit found that in one logged sequence, the AI made seven channel search calls and only surfaced an archived private channel from 2019, never locating the intended target.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Blockmachine Aims to Disrupt Ethereum RPC Market with Decentralized Node Marketplace

Ethereum applications rely on RPC providers to interact with the blockchain without running their own nodes, a role currently dominated by centralized services like Alchemy, Infura, and QuickNode. Critics argue these providers suffer from opacity and oligopoly-driven pricing, as users must trust the provider's own reporting on data accuracy and billing. Blockmachine is a newer entrant that replaces the single-operator model with a competitive marketplace where independent node operators bid on each request. The platform claims to enforce data correctness through real-time cryptographic verification, rejecting responses that fail proof checks and removing dishonest nodes from the network. Its stated goal is to address all four key pressures on RPC infrastructure simultaneously: correctness, low latency, full method coverage, and transparent pricing.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

memlineage v0.1.0 offers two-layer defense against memory poisoning in LLM agents

Developers have released memlineage v0.1.0, a Python library designed to protect AI agents with persistent memory from adversarial memory poisoning attacks that operate through normal interaction channels. The tool combines cryptographic provenance using Ed25519 signatures with behavioral detection based on tool-call trajectory analysis, addressing weaknesses that recent academic research identified in content-filter and trust-score-based defenses. Each stored memory is signed at write time and tied to a verified principal, while a derivation graph enforces that any memory traceable to an untrusted source can never be laundered to a trusted tier. A SensitiveActionGate blocks irreversible actions whose justification traces back to an untrusted ancestor, achieving an attack success rate of 0.000 against Layer A and an AUC of 0.978 with attacker camouflage on Layer B. The release was verified by an independent auditor who ran custom adversarial tests, with the behavioral detection layer currently set to alert-only mode pending false-positive rate measurement.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Multi-Stage Docker Builds Cut Python App Image Size from 1.2GB to 85MB

A software team discovered their Docker image for a simple Python web app had ballooned to 1.2GB, causing CI pipeline runs to take over 11 minutes. The bloat stemmed from common issues such as build dependencies leaking into runtime, uncleaned package manager caches, and the additive nature of Docker layers. By adopting multi-stage builds, the team used a separate builder stage to compile dependencies and then copied only the necessary artifacts into a slim runtime image. Switching the base image from python:3.11 to python:3.11-slim immediately reduced the image to around 210MB, and moving to an Alpine-based image brought it down further to approximately 85MB. The approach keeps heavy build tools like gcc and g++ out of the final image, significantly speeding up deployments without changing application functionality.

0 comments Read more at DEV Community