CK Context & knowledge
Context engineering, code knowledge graphs, retrieval over code, agent memory.
From the issues
Stop Trimming Your Agent's Prompt: why fewer tokens can mean more compute
Yes, trimming agent history can cost more. In The Physical Cost of Context, keeping only the last five tool outputs cut prompt volume 33.3% but raised uncached prefill 608.5%, because every edit broke the prompt cache. Keep history append-only, prune once at task boundaries, and measure cache misses, not prompt tokens.
One Skill Too Many: the lookalike that quietly took the job
Check which agent skill actually ran, not which one you installed. A study of 20,947 repositories found nearly one in four installed skills duplicates another. In 312 confirmed pairs, the lookalike took over in one in five runs, and only 0.9% of those runs named the skill used.
"Up to 45x fewer retries"
A launch number with no method is a ceiling, not a result. GitLab says its Orbit code graph gives agents up to 45x fewer retries and 4.5x fewer tokens, self-reported, with 3,500+ organisations in beta. The release names no task, baseline or median. Ask for the floor before you clap for the ceiling.
Pack a Tiffin, Not the Kitchen: your context file is billing you every turn
A long AGENTS.md doesn't make your coding agent better. ETH Zurich found context files raised inference cost by over 20% with no general gain in task success, and repository overviews didn't help at all. Keep only the rules the code can't tell the agent. Delete the directory tour.
Numbers on this track
| Number | What it measures | Source | Checked | Track · issue |
|---|---|---|---|---|
| >20% | Extra inference cost from repository context files (AGENTS.md), with no general gain in task success | Gloaguen et al., ETH Zurich, arXiv:2602.11988 | ✓verified at source | CKContext & knowledge Issue #01 |
| ~66% | Share of Cursor's system prompt trimmed | Cursor blog | self-reported | CKContext & knowledge Issue #01 |
| 7% | Cursor's total token reduction, "without degrading quality" (method not shown) | Cursor blog | self-reported | CKContext & knowledge Issue #01 |
| 22–25 of 27 | Top-four model scores on enterprise questions where the rules are stated | Era by Eon, arXiv:2609.30055 | ✓verified at source | CKContext & knowledge Issue #01 |
| ≤6 of 24 | Scores of four of six models when the answer hinges on a hidden fact in contradictory records | Era by Eon, arXiv:2609.30055 | ✓verified at source | CKContext & knowledge Issue #01 |
| ~1 in 4 | Installed agent skills that sit beside another skill doing the same job (20,947 repositories) | Yan et al., One Skill Too Many, arXiv:2610.11647 | ✓verified at source | CKContext & knowledge Issue #02 |
| ~1 in 5 | Runs in which the lookalike skill took over, task completion unchanged (312 pairs, 6,368 runs) | Yan et al., One Skill Too Many, arXiv:2610.11647 | ✓verified at source | CKContext & knowledge Issue #02 |
| 0.9% | Substituted runs whose final reply named the skill actually used | Yan et al., One Skill Too Many, arXiv:2610.11647 | ✓verified at source | CKContext & knowledge Issue #02 |
| up to 45x | GitLab Orbit's claimed reduction in agent retries (no method given) | GitLab press release | self-reported | CKContext & knowledge Issue #02 |
| up to 4.5x | GitLab Orbit's claimed reduction in tokens (no method given) | GitLab press release | self-reported | CKContext & knowledge Issue #02 |
| 3,500+ | Organisations in the Orbit beta, per GitLab | GitLab press release | self-reported | CKContext & knowledge Issue #02 |
| 33.3% | Cut in prompt volume from keeping only the last five tool observations (Mask-5) | Yu, Zhao et al., arXiv:2610.34763 | ✓verified at source | CKContext & knowledge Issue #04 |
| +608.5% | Rise in uncached prefill under Mask-5 (energy +125.6%; failed the quality gate, 26 against 30 of 100 resolved). One GPU, one request at a time | Yu, Zhao et al., arXiv:2610.34763 | ✓verified at source | CKContext & knowledge Issue #04 |
| 35.7% / 33.2% | Cut in model calls and in energy from preloading 10 retrieved code blocks (Eager), with no quality loss | Yu, Zhao et al., arXiv:2610.34763 | ✓verified at source | CKContext & knowledge Issue #04 |
| up to 7x | Claimed extra cost of later tokens in long contexts (abstract) | NVIDIA, arXiv:2611.05306 | ✓verified at source | CKContext & knowledge Issue #04 |
Papers on this track
- Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? arXiv:2602.11988
- Era by Eon: Benchmarking Enterprise Agents on Hidden Knowledge arXiv:2609.30055
- One Skill Too Many: How Co-Installed Skills Conflict in Coding Agents arXiv:2610.11647
- The Physical Cost of Context arXiv:2610.34763
- You don't need long context arXiv:2611.05306
Other tracks
ISIntent & specs ABAgents that build UMUnderstanding & modernisation VTVerification & trust POPeople & operating model
Last updated . Reuse with credit under CC BY 4.0: “The Dabbawala Protocol, thedabbawalaprotocol.com”.