Daybook, 14 July 2026

Evaluating a small agent — and watching the scope widen. Most of the day went into extensive evaluations and development of a small agent that writes scripts, workflows, and analytics-engineering tasks. And the scope keeps growing — not just AI for managers, but for field operations, delivery, and even engineering work. The bigger target is still the platform: people working on the data and the context, and building agents that can be infused into other workflows.

Tried GEPA today. Used GEPA — the reflective, Pareto-based prompt optimizer from DSPy — to optimize a few of the agents. In practice it's really a hill-climbing task on the agent harness and the tool curation. And it's the honest first move inside an enterprise: optimize the harness and the prompts before you reach for anything heavier. It was fun.

The post-training question I'm actually chewing on. Not every enterprise can — or should — invest in post-training a model. One, it needs infrastructure; two, most enterprises have many diffuse use cases rather than one sharp capability. Post-training really earns its keep at a company like Harvey — law is pinpointed, you can curate all the tasks into a very good benchmark, and the post-training research can push that product's Pareto frontier. But an enterprise with operations, domain after domain — it has to prioritize where the training investment actually returns value. So it's not just “throw RL at it,” or randomly throwing GRPO at things. In this setting you start with GEPA-style prompt and harness optimization, and only climb toward agentic RL and self-improvement where a domain is sharp enough to be worth it. It's debatable, and that's the part I'm sitting in.

Self-improvement, when the weights are frozen. Most real deployments can't touch the weights — you're on a hosted or shared open-weights model, not something you own on-prem. So continual learning and adaptive context have to live outside the model, captured somewhere the system carries forward. Call it memetic: the unit that survives and replicates isn't a weight update, it's a captured pattern — a markdown file, or a graph. It rhymes with the context-graph work from earlier this week: the substrate holds what's learned.

A question that kept circling. Agent-to-agent communication is drifting toward opaque channels — hidden states, KV-cache — chased for raw speed. I doubt that trade is right: communication should stay legible, readable by humans and agents alike, because that legibility is the ground guardrails, oversight, and prompt-injection defense stand on. The fastest channel may quietly be the least safe one. Worth digging into properly — a couple of references at the bottom.

Reading — the hierarchy of work. In The Meaning of Your Life, there's this pyramid: the base is necessity — food, shelter, survival. It rises through work-as-duty and service to others, and at the top, work-as-calling — where it becomes a source of meaning rather than just a means of living. Franklin, Mother Teresa, Marx — each treating their work as a calling. Reading it chapter by chapter, I keep taking my own inventory: some of this I'm already living, some is worth moving toward. The turn into mysticism and spirituality isn't about practice for me, but it's good to read, and it clarifies.

And on the research-engineering vs forward-deployed map from earlier this week: honestly, I think I can do both. The open question is just which one stays interesting.

A frame that clicked while reading. Research engineering is really about moving the frontier — but the skill isn't “optimize everything.” It's multi-objective, and multi-objective means picking a point on a Pareto trade-off, not maximizing cost, capability, and latency all at once. So the sharper questions are: which axis am I moving; for whose frontier — the field's (that's a lab's job) or this company's internal one (an enterprise mostly reaching an existing frontier, cheaper); and with a lever whose cost the objective actually justifies. That last one is the post-training test — reach for GRPO / agentic RL only where the cheaper levers are exhausted and the reward is checkable. And it's good to see it line up with what Exp-2 already found from the job descriptions: capability is table stakes, and depth on one axis beats broad coverage.

References I'm sitting with — on why agent communication should stay legible: Secret Collusion among AI Agents (NeurIPS 2024) · Latent attacks in latent-based multi-agent systems.

Settled · 14 Jul 2026