IF-MAP, LangChain retrieval patterns, Baseten RL environments, and MCP gateways all attack the same cost: doing everything, everywhere

Optimize only the agents that matter

A preprint on multi-agent systems starts from an uncomfortable fact: improving every agent in an LLM-based multi-agent system can be prohibitively costly, and failure attribution does not directly reveal the system-level gains achievable by improving an agent. The authors propose IF-MAP, which identifies high-impact agents in multi-agent LLM systems via prompt perturbation analysis. It measures end-to-end loss changes under controlled semantic prompt perturbations, fits a regularized local response model in a low-dimensional prompt-embedding subspace, and uses the estimated response norm as a priority score. It is a black-box procedure requiring no access to model parameters or gradients, which means it works on systems you do not own end to end. m:…7038

The preprint's results reward the prioritization. Optimizing selected agents' prompts achieves a macro-average gain of 3.81 percentage points, against 1.85 percentage points for uniform agent selection. Upgrading selected agents yields macro-average gains of 7.62 to 16.42 percentage points, against 3.81 to 8.62 for uniform upgrades. The token savings are the headline for anyone paying the bill: IF-MAP reduces Qwen3-8B token use by 74.2% relative to optimizing every agent, reduces DeepSeek-V3.2 token use by 35.2%, and reduces target-model tokens by 63.6 to 64.9% relative to upgrading every agent. The study evaluated six multi-agent workflows spanning code generation, question answering, and mathematical reasoning. The lesson generalizes: in a multi-agent system, the budget question comes before the capability question. m:…7038

Retrieve, offload, delegate

LangChain's post on context engineering makes the same argument at the level of a single agent's context window. Agents have limited context windows, the post says, which makes retrieving information necessary for task completion, and inefficient searching consumes tokens and time, potentially filling the context window with failed search results. Dumping full data exports or multiple documents into the context window can leave no room for the actual task, and irrelevant or stale context can lead agents to build on incorrect information, propagating mistakes. The failure mode is mundane and expensive: the agent spends its budget reading, not doing. m:…3700

The post's pattern is named retrieve, offload, delegate. The Deep Agent harness uses glob to find files by name pattern and grep to search inside files without loading them into the context window. The pattern involves saving search results to files and returning only the file paths to the agent, and subagents can be used to filter retrieved files and return only key facts to the main agent. The harness automatically offloads tool results exceeding 20,000 tokens to files, and its read_file tool reads files one page at a time, defaulting to 100 lines. Retrieval tools backed by vector stores allow for semantic search by meaning rather than exact keywords, and can interface with SQL databases, search APIs, or MCP servers. Agentic retrieval lets agents decide when to search, what to search for, and whether additional searches are needed. Search know-how can be packaged as a skill, which loads full instructions only when called. m:…3700

One warning in the post deserves emphasis: retrieved text should be treated as untrusted because it may contain prompt injections. Offloading and delegation do not remove that risk; they move it. A subagent that filters files is also a surface for injected instructions. The architecture saves tokens, but the trust boundary still has to be designed, not assumed. m:…3700

Post-training and data access get the same treatment

Baseten applies the prioritization logic to post-training. The company is introducing the Baseten Environment Network, a group of vetted RL environment companies. Baseten will scope engagements, while partners work directly with customers to build environments, tasks, and rubrics, and Baseten retains responsibility for training and deploying models onto its production inference infrastructure. The company has previously built custom RL environments for clients including Harvey, Lovable, and Gamma, and says customers can use the custom RL environment systems built by the network in perpetuity. The company plans to open-source a set of frontier-quality environments and its auditing systems. The pitch is that post-training data systems are becoming something you own, built by specialists, rather than something you rent by the prompt. m:…6291

DatI closes the loop on the data side. The company describes DatI as a lightweight semantic gateway connecting AI agents to enterprise databases, supporting publishing databases as MCP services using Streamable HTTP. The latest update introduces schema-level fine-grained isolation and permission control, so users can publish MCP services by specifying a schema instead of manually configuring individual tables. DatI now allows AI agents to create intermediate or analysis tables within authorized schemas, automatically registers metadata and semantic indices for newly created tables, and supports MySQL, PostgreSQL, ClickHouse, and Doris databases. The pattern is the same as the rest of the story: give the agent a bounded, permissioned surface instead of the whole database, and let it work inside that boundary. m:…8370