IBM Research: General Agent Evaluation (Exgentic Framework)

IBM Research released CUGA / Shared entity: IBM Research / Same source domain / What happened next

IBM Research ships CUGA: 24 copyable examples for agentic apps on a lightweight harness.

Linked by a graph relationship (IBM Research released CUGA); both cover IBM Research; reported by the same outlet (arxiv.org).

Same source domain / Shared topic / Tension

279-Paper Review Says AI Coding Benchmarks Overstate Skill

Reported by the same outlet (arxiv.org); overlapping topics (agent, evaluation, finding); pushes against this story (vs).

Even top models miss errors in over half of buggy code submissions.

Reported by the same outlet (arxiv.org); overlapping topics (agent, evaluation, model); pushes against this story (but).

Run a dedicated spec-elicitation agent before the coding agent.

Reported by the same outlet (arxiv.org); overlapping topics (agent, model); pushes against this story (against).

Gate on metadata before the agent acts, because no model tier catches stale-but-well-formed data.

Reported by the same outlet (arxiv.org); overlapping topics (agent, model); pushes against this story (but).

Same source domain / Shared topic / Downstream implication

Seven open-weight models fail cooperation identically: they pay the query cost without transmitting the information.

Reported by the same outlet (arxiv.org); overlapping topics (agent, model); traces where this leads (downstream).

Same source domain / Shared topic / Tension

Frontier models agree on financial decisions 95% of the time but follow the same tool path only 77%.

Reported by the same outlet (arxiv.org); overlapping topics (agent, model); pushes against this story (but).

66.5% of Malicious GitHub Issues Get Past Every Guardrail in Cursor, Claude Code, and Codex

Reported by the same outlet (arxiv.org); overlapping topics (agent, model); pushes against this story (against).