The agent memory shift, minimal harnesses, and sampling drift

00:00:04 DAIR.AI reposted a Salesforce AI Research paper on agent memory today. The finding is straightforward: store raw trajectories and decide what to extract from them when the next task arrives, rather than summarizing each run the moment it ends. Just-in-Time Memory uses a curator that reads stored trajectories on demand.

00:00:26 You trade upfront summary fidelity for retrieval quality later — a real decision with real costs. If your retrieval layer is slow or inaccurate, you lose the benefit. If your storage scales linearly with agent runtime, you pay for it in infrastructure. Summarization still operates as the default compression strategy for agent history because it's convenient: you run something, you get a summary, you move on.

00:00:55 But the paper suggests summaries throw away information that could matter for entirely different downstream tasks. If the next task needs the exact tool arguments from three turns ago, not the gist of what happened, you've already compressed past the useful signal.

00:01:13 Hamel Husain's guide on synthetic data generation hits the same note: define what you need to test first, generate combinations of those parameters, then run them through your application. The operational advice always runs backward-looking — start from what you need, not from what the system naturally produces.

00:01:35 MIT CSAIL released a paper today introducing JAZ, an agent framework with exactly one primitive: invoke. The large language model writes code directly. There's no routing layer, no tool registry, no abstraction hierarchy. Omar Kholeif flagged it on X with the simple takeaway you'd expect from someone who's seen enough harnesses to know better: keep them minimal.

00:02:01 Minimalism in infrastructure usually wins because complexity has its own inertia. Every added primitive — a new routing pattern, a new tool-calling convention, a new profile layer — becomes something you have to document, test, and debug. The MIT team found that the simplest possible dispatch mechanism, paired with a model capable of writing executable code on the fly, outperformed more structured approaches on self-improvement tasks.

00:02:32 Agent frameworks proliferated in 2025 because every team had a different tool ecosystem and a different opinion on what abstraction was needed. As models get better at code generation and structured output, the framework layer starts looking less like a solution and more like overhead.

00:02:52 The practical choice isn't which harness to pick — it's whether you need one at all.

00:02:58 Roland Gavrilescu, who left XAI to build introspection tools with his co-founder, published a talk today proposing recipes as a product category for agent configuration. His framing is useful: an agent recipe is a portable, versioned git repo encoding evals, harness profiles, human judgment, and the feedback loops between them.

00:03:22 Not code. Configuration. The stuff you tune after the model stops moving the needle. He calls the core operation system distillation — understanding what went well and wrong in one loop, processing that into a second loop, iterating. You treat it like RL data recipes: track the recipe, change it continuously to combat hallucination or reward hacking, and build toward a final stack.

00:03:49 The current version applies this to harnesses and agent profiles rather than training data. What's interesting about Gavrilescu's pitch is that he's productizing something most teams currently keep in Slack threads and Google Docs. If recipes become a real format — versioned, distributable, testable — then agent tuning looks a lot more like the ML ops pipeline it should have always been.

00:04:17 You inherit other people's taste through their recipes.

00:04:21 Lasso Security tested Anthropic's SynthID-Text watermarking across six models and found it changes tool-call correctness in ways that are hard to see in aggregate. The mechanism is straightforward: tournament sampling adds a random seed generator, sampling algorithm, and scoring function to token selection.

00:04:43 Non-distortionary in the sense that the distribution stays the same on average over many runs. But for any single run with a fixed key, token selection shifts. And when those shifted tokens land on agent-relevant fields — tool names, argument values, boolean flags — they change what the agent actually does.

00:05:05 Across 21 model-temperature combinations, paired disagreement averages 6.5%. For phi-4 at temperature one, sixteen point eight percent of tool-call verdicts differ between watermarked and unwatermarked runs while net accuracy loss is only two point eighty-seven points.

00:05:24 The churn exceeds what a temperature change produces on some models. The effect is key-dependent and model-dependent. Lasso tested multiple keys rather than relying on one. What they report are both net performance changes and paired disagreement — the former tells you the average, the latter tells you how many individual calls flip.

00:05:47 Anthropic deploys watermarking at the model level. If you're building an agent on Claude through the API, the watermark affects your tool calling even if your agent application doesn't know it's there. This isn't a prompt injection exploit — it's what happens when invisible provenance markers change which path through the token space gets sampled.

00:06:12 Arthur Mensch, Mistral AI's co-founder, gave an interview to Le Monde today arguing that American tech giants use risk rhetoric to control and close off the market — that warnings about AI-pocalypse scenarios serve competitive closure rather than genuine safety.

00:06:31 Mistral raised three billion euros in September. A new model launches "in the coming weeks." Mensch is pushing for state guarantees to fund European data centers. The position is familiar: open, accessible AI as a counter to American concentration. What makes it different this time is the capital behind it.

00:06:53 Three billion euros isn't a statement of faith — it's a commitment to building infrastructure that can compete. Whether Mensch's argument holds up will depend on whether Mistral's next model closes gaps with frontier US models or if the competitive position remains in rhetoric and fundraising rather than capability.

00:07:16 Control shows up across these stories. Not the abstract kind — the concrete kind that shows up as a design choice in agent architecture. Store raw trajectories instead of summarizing: you control what information survives compression. Use one primitive for dispatch: you control how much abstraction gets between the model and its tools.

00:07:37 Version harness configs as recipes: you control who inherits your tuning decisions. And now even watermarking, designed for provenance, introduces behavioral churn into agent tool calling. All of these are about where decisions get made and what information gets preserved when systems scale beyond single conversations.

00:07:58 The operational detail that stuck with me is Lasso's paired disagreement metric — six and a half percent average churn masked by flat aggregate scores. That's the kind of number you only notice if you look at individual call pairs instead of overall accuracy.