Agent shorts: ToolSpace, Axion 1.0, DEPTH, Switch-SWE, Caper, Recursa, GUI-agent RL, DeepSeek's super-agent harness

Sources: HuggingFace 2026-11-04: ToolSpace 2611.05036, Axion 1.0 2611.03233, DEPTH 2611.01927, Switch-SWE 2611.02017, Caper 2610.40753, Recursa 2611.04035, GUI agents: where to learn 2610.39251, EvoTeam 2610.39813 · X: RSI-Dojo 2611.03698 via @dair_ai · Ken Huang: DeepSeek super-agent, Harness Engineering book · Agent Arena (@arena)

TL;DR. The agent papers this window are mostly about measuring and building better environments, plus two cost-efficiency results. ToolSpace mines real MCP execution traces for data, state and retry dependencies between adjacent tool calls (deterministic rules, no LLM judge), clusters them into environments, and decompiles state-aware random walks into tasks with assertion checks; "Pro" versions of AutomationBench and tau3-Bench Banking drop seven frontier models by up to 38.2 points. DEPTH shows synthetic terminal tasks add work, not depth: 94% of solved synthetic tasks pass within three turns, and depth (turns that must happen in sequence) predicts difficulty beyond work. Switch-SWE: interleaving unfinished tasks in the same repo costs seven coding agents 11.5 points on average (6-15). Caper pre-executes each shell command in a snapshotted sandbox and shows an independent approver the observed effects; harmful-outcome rate 6.7% vs at least 14.8% for six existing approvers, 7.8% vs 35.3% when the harm hides inside invoked tools, while blocking only 1% of safe twins. Axion 1.0 is a fully open agent model trained in only SFT, SFT, RL (no mid-training, DPO or OPD), with the whole recipe released. GUI-agent RL that restarts from saved intermediate checkpoints instead of the task root cuts environment interaction 56-61% versus vanilla GRPO while improving success. Recursa grows a skill-and-memory repository with parallel learners and raises Harvey LAB held-out all-pass from 30% to 55% on DeepSeek-V4.1-Flash. RSI-Dojo turns papers into 281 post-training tasks; fine-tuning on filtered trajectories lifts a 35B-A3B MoE by 16.87 points on PostTrainBench.

Harness practice

DeepSeek's super-agent (dsh): an event-driven daemon over Git, one worktree per task, deterministic admission gates before any inference, and an orchestrator that never runs shell commands or waits on workers (a router on a lighter model). Ken Huang extracts it into a cross-harness skill with a "Rule of Three" ceiling on parallel subagents.
Ken Huang's book names ten harness pattern families (autonomy boundaries, scoped identity, intent contracts, context isolation, memory governance, observability, runtime steering, fleet scalability, data governance, empirical evaluation). A taxonomy, not evidence.
Agent Arena economics: an hour of agent work averages 332 tool calls; median cost per hour $10.51 (Gemini 4 Carbon), $21.13 (Opus 5.5), $31.08 (GPT-6 Astra). On the costliest 5% of sessions OpenAI and Anthropic hold 7 of 11 frontier spots versus 10 of 11 on cheap tasks.

How it relates to prior wiki pages

DEPTH and ToolSpace confirm the Prime Intellect benchmark audit (11-04: graders that reject correct answers, leaked solutions): agent benchmarks overstate difficulty and progress. Both replace LLM-judged difficulty with execution-grounded structure.
Caper extends LeanGuard (11-04, tool calls checked against a Lean4 policy that fails closed): approval is moving from reading the command to observing its effects.
Switch-SWE pairs with the 11-04 continual-harness-evolution result (94% to 32% on SWE-Bench after evolving elsewhere): both measure interference across tasks, at runtime and at evolution time.
GUI-agent checkpoint restarts are the training-side cousin of NVIDIA's compaction (today): don't replay a prefix you have already mastered.