Research

Not vibes — implemented papers.

FORKSCAPE's design follows published research. Every card links to the paper.

The problem. −39%

LLMs Get Lost in Multi-Turn Conversation

Model quality drops 39% on average as a linear chat grows. The longer the thread, the worse every answer gets.

Microsoft Research · arxiv.org/abs/2505.06120 ↗
The fix — FORKSCAPE is this, productized. −58%

Context Branching for LLM Conversations

Branching cuts context size 58.1% and improves response quality. Isolated branches beat one growing transcript.

arXiv · arxiv.org/abs/2512.13914 ↗

Six papers, shipped as features.

CodeAct — executable code actions

Agents act by writing and running code instead of emitting brittle tool schemas. Shipped as our Code Agent mode.

arxiv.org/abs/2402.01030 ↗

Parsimonious delegation — Uno-Orchestra

Cost-aware triage before any fan-out: wasteful multi-agent runs are refused up front, not billed and regretted.

arxiv.org/abs/2605.05007 ↗

Lineage replay

Unchanged re-runs replay from cache and cost 0 tokens. Only edited nodes — and what depends on them — run live.

arxiv.org/abs/2605.06365 ↗

ACE — agentic context engineering

Build insights are distilled into durable project knowledge, so every canvas starts smarter than the last.

arxiv.org/abs/2510.04618 ↗

TRACE — correction distillation

Your corrections become enforced rules. Fix an agent once and the fix persists across future runs.

arxiv.org/abs/2606.13174 ↗

SETS — rubric verification

A post-run judge scores deliverables against a rubric, with one bounded revision round — quality without infinite loops.

arxiv.org/abs/2501.19306 ↗