CliffCompaction cuts coding-agent costs by throwing old context away

Good morning 👋 A new paper says the fix for expensive long-horizon coding agents isn't a smarter memory strategy. It's a dumber one.

In today's issue:

• 🔭 CliffCompaction cuts coding-agent costs by throwing old context away
• 🔬 A new MCP supply-chain attack, and hidden reasoning pulled into view
• 🛠️ Put CliffCompaction's own trick to work on your agent this week
• 🔗 OpenAI's own third-party audit rulebook, and a push to make benchmark scores reproducible

🔭 THE ONE THING

🧮 CliffCompaction cuts coding-agent costs by throwing old context away

CliffCompaction cuts the cost of long-horizon coding agents by up to 50%, and it does it the boring way: truncate or drop old context instead of paying to rephrase it, so nothing drifts across sessions. On KernelBench that discipline bought a 3.58x CUDA kernel speedup after 400 steps, and under parallel test-time scaling it let Kimi K2.6 match Opus 4.7 at a fraction of the spend. The unglamorous part is the point: every other approach to long-horizon cost has tried to get smarter about summarizing history, and this one just says stop paying for most of it. It's open source and already wired for Claude Code, so there's no reason to wait for someone else's benchmark to tell you whether it holds up on your own agent. Paper

🔬 RESEARCH HIGHLIGHTS

• A2M treats MCP tool selection like a software supply chain, because that's what it is. The attack poisons a tool's metadata to get it picked, then poisons its output to steer the agent, and hit a 93.6% malicious invocation rate on GLM-4.6. It ports to other models with zero retraining, at a 63.6% hit rate. If your agent pulls tools from third-party MCP servers on trust, this is the paper that says stop. Paper

• Hidden chain-of-thought, extracted. Register a custom tool through a standard API feature, and closed models will externalize reasoning they were supposed to keep hidden, GPT-6 Astra included. The pulled-out traces perform as well as the model's native reasoning. The finding worth sitting with: Astra reasons tight, locking onto the right path early and only surfacing the steps that matter, which tells you more about how it thinks than another benchmark table would. Paper

🛠️ TRY THIS

Put CliffCompaction's own trick to work on your own agent

1. Pick a long-running coding-agent workflow where the token bill climbs the longer a session runs.

2. Stop summarizing or rephrasing old context. Hard-drop anything past a fixed turn count instead, same as CliffCompaction does.

3. Run a week of your usual tasks against both versions, full-context and truncated, and log token spend and task success for each.

4. If success holds and spend drops, make the truncation the default. If it doesn't, you've learned the history was actually load-bearing.

Prompt: Find where this agent keeps full conversation history across steps, and rewrite it to hard-drop context older than N turns instead of summarizing it.

Worth a look

• AWS Strands Evals now scores whether an agent picked the right skill and actually followed it, not just whether the final answer read fine. Useful if your agent has more than one skill to choose from. Evaluate skill-equipped agents with Strands Evals
• Hugging Face Tokenizers hit v1 with real encode, decode, and scaling benchmarks. Worth a skim before you pick a tokenizer for a new pipeline. tokenizers v1

🔗 QUICK LINKS

• OpenAI published its own rulebook for who gets to independently audit its frontier models, before anyone else could write the rules for them. link
• Hugging Face and the UK AI Safety Institute are building a shared methodology so AI benchmark scores can actually be reproduced by someone other than the vendor. link

That's the lineup. Go find out how much of your agent's context you can actually afford to lose.

Pradeep Perugu