Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-31 · high
It isolates the exact agent and step behind a failure, then reflects only there, gaining 22 to 27 points over three prior baselines.
Why now: DoCtOR's numbers are new to the record as of August 31, 2026, and they give multi-agent builders a measured comparison against Reflexion, Retroformer and COPPER.
DoCtOR pinpoints the single agent and step that caused a multi-agent run to fail, then rewrites only that agent's next move.
Most multi-agent systems do the opposite. A failure triggers a reflection prompt for every agent in the run, including ones that behaved correctly. That waste is specific. DoCtOR's targeted method gains 22 points on HotPotQA, 26 on ChartQAPro and 27 on Mind2Web over initial success rates, beating Reflexion, Retroformer and COPPER.
The method runs automated failure attribution first, then uses counterfactual reasoning to work out what the failing step should have produced instead. Only the agent responsible for that step gets the reflection prompt. Every other agent's memory stays untouched.
A smaller result in the paper holds up on its own. In low-resource settings, reflecting only on the steps after the decisive error matched reflecting on the whole trajectory. The waste this fixes is one I've caused myself. Broadcast reflection writes a wrong lesson into the memory of an agent that did nothing wrong.
Each link below shares sources, entities, or timing with this story.
arXiv 2608.05906 keeps a dual-polarity memory of verified corrections and observed dead ends for Text-to-SQL repair: 66.34% to 69.79% on Spider, 47.35% to 48.44% on BIRD. Then the authors say the quiet part: paired analysis supports the Spider gain but is weak on BIRD, MERIT i...
Skill-α (arXiv 2608.01678) reframes skill generation as RL over sequential edits, decomposing skill construction into individually evaluable changes. The novel signal is a rollback reward that scores each modification by comparing downstream task execution using the original s...
Microsoft's July 23 release targets a genuine gap: harness-based agents like Claude Code and Codex drive multi-turn reasoning, tool use, and external system access but were hard to train end-to-end with standard open RL infrastructure. The trick is decoupling training from inf...
If you're tuning an agent system, upgrade the model doing the tuning before you rewrite a single line of the harness. That's the finding from HarnessOpt-Bench (arXiv 2608.06301, Scale AI), which tests whether frontier models can improve an agent *system* rather than write code...
For two years the technique was accumulation. Longer system prompts, longer CLAUDE.md, more numbered do/don't lists, more "always verify your work" imperatives. Anthropic's context-engineering guidance for Claude 5 models inverts it, with an 80% deletion figure attached. The s...
Recuris (arXiv 2608.24876) keeps a Working Memory tracking current task progress separate from an Experiential Memory of learned skills, so skill selection indexes against what the task needs now rather than the whole history. It improves 35 of 37 model-benchmark pairs, gains...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.