AI Agents
Autonomous agents, multi-agent workflows and orchestration.
54 published articles
Coding agents · Devin's two-model harness
Paying 2x more per token made Devin's Fusion 9% cheaper
Devin's Fusion runs a frontier lead model beside a cheaper sidekick, each with its own context. Devin's benchmark tables show task costs dropping as much as 46%, and one case where a lead costing twice as much per token produced a 9% cheaper session.
2026-09-26
Artificial intelligence · Agent safety
SafeEvolve cuts agent attack success 3x by evolving harness and model together
SafeEvolve, an arXiv paper submitted on 2 September, pairs runtime harness updates with policy training so an agent's own trajectories drive both. It reports a threefold cut in attack success on AgentDojo and a modest utility gain.
2026-09-25
Autonomous coding agents · arXiv preprint
Harness-of-Harness claims a 52.25% lift by running other coding agents
HoH wraps existing coding-agent harnesses in repeated plan-code-test loops and reports an average 52.25% relative gain across three benchmark suites. The abstract leaves the underlying metric, the resource cost and the per-model split undefined.
2026-09-25
AI alignment research on arXiv, 10 September 2026
Giving AI an "artificial id" could fix control, or entrench failure
The preprint argues alignment should be a property of a continuing agentic system rather than a single model response, and that per-response checks cannot remove a bad strategy which survives into the next task. Its evidence is a controller too small to reason.
2026-09-21
AI research: multi-agent decision-making on arXiv
A panel of AI agents can agree and still be wrong, paper argues
Voting, electoral rules and LLM judges all map evidence to labels in one direction, so their errors can be correlated. A new paper proposes ranking agents by a reverse Bayesian posterior instead, and reports its biggest gains where agents disagree.
2026-09-19
AI research: an arXiv preprint dated 10 September 2026
COBRA-Skills cuts agent skill tuning costs by 55-58% on 50 examples
COBRA-Skills pairs contextual-bandit prioritization with skill evolution to cut agent skill optimization costs by 55-58% against SkillOpt, using 50 unique examples per benchmark.
2026-09-19
Agent post-training, arXiv preprint, 3 Sept 2026
Terminal-Universe mines 37,300 training rooms from old agent runs
Terminal-Universe replays recorded file operations to rebuild the workspaces coding agents once ran in, then generates fresh tasks from them. The preprint claims 37,300 environments and double-digit benchmark gains, all self-reported.
2026-09-16
Multi-agent AI systems and emergent behavior
Cheating spread through a 100-agent AI swarm. So did the whistleblowers
A new arXiv case study follows 100 autonomous LLM agents proving math conjectures, one of them finding an evaluation exploit, and a separate cohort organizing against the fraud. No human intervened.
2026-09-16
AI research automation, arXiv, September 2026
DisCo reports a 134.3% MLE-bench gain with its backbone held fixed
DisCo distills 1,000 ML repositories into more than 5,000 reusable skills, and reports large benchmark gains with the model, harness and execution budget held fixed. The figures are self-reported and unreplicated.
2026-09-16
Design Tools
Open Design 0.18.0 ends the 'is this the latest version?' meeting
Open Design 0.18.0 ships a shared Team workspace with live read-only mirrors and lets Codex call the design engine headlessly. 115 pull requests from 22 contributors landed in two days. Here's what changed and the agent-tooling trend it fits.
2026-08-21
Agent Memory
TEPA: for AI agents, a stale memory is worse than none
Agent memory has a falsifiability problem, a new arXiv preprint argues: stale facts stay retrievable and pollute the prompt. Its TEPA mechanism revokes superseded memories, and in drift tests naive memory scored below no memory (0.210 vs 0.309) while TEPA reached 0.950.
2026-08-19
AI Agents
PsychoAgent gave AI agents emotional memory. The raters found no edge
PsychoAgent gives LLM agents a separate affective memory so emotionally salient, conflict-laden traces can outweigh merely topical ones. In three conflict scenarios it retrieved more conflict-critical memories than both baselines (0.933 vs 0.500 and 0.667), yet five blinded raters found no significant quality edge.
2026-08-19