Agents get their Heroku moment
-
Nº XVIII
- Date
- 19 May 2026
- Issue
- 18
- Stories
- Six
- Editor
- ARC
Today: backend plumbing for agents gets serious, energy-based models learn physics, and an MCP server finally makes life-science KGs talk.
InsForge pitches Heroku for agents
InsForge open-sourced a backend platform built specifically for coding agents — auth, databases, storage, and deployment surfaced as MCP tools (Model Context Protocol, Anthropic's spec for letting agents talk to services) rather than dashboards meant for humans. The pitch: agents shouldn't have to click through Heroku UIs they can't see. By exposing infrastructure primitives as agent-native calls, InsForge raises the floor for what an autonomous coding agent can actually ship end-to-end, including the bio-tooling agents that increasingly need their own persistent state, secrets, and storage to run multi-day discovery loops.
Energy models enforce molecular physics
Energy-based generative models now produce molecules that respect physical constraints — bond lengths, angles, and conformer energies — rather than the geometrically plausible but physically nonsense outputs diffusion models often emit. The result moves generative chemistry one step closer to candidates that don't need a downstream filter to throw out half the proposals, tightening the link between in-silico design and what actually docks.
SkillGenBench scores agent skill libraries
SkillGenBench evaluates how well LLM agents generate and reuse their own skills — the callable analysis routines agents accumulate over a session. Most agent frameworks claim skill libraries; few measure whether the skills are actually reusable. The benchmark anchors a reference point for skill-quality claims, relevant to any agent platform pitching biology workflows where the same analysis pattern recurs across experiments.
TogoMCP queries life-science KGs in English
TogoMCP wires life-science knowledge graphs to LLMs via MCP and schema-guided prompting, letting researchers ask in plain English instead of writing SPARQL. Closes one of the longest-standing usability gaps in bio knowledge graphs — the data has always been there; the query language was the barrier.
Sequence pretraining finds cryptic pockets
Sequence-only pretraining for drug-target binding sites now detects cryptic pockets and predicts binding kinetics, not just affinity. Pushes structure-free methods into territory that previously demanded MD simulation, lowering the compute bar for early-stage target assessment.
Anthropic buys Stainless
Anthropic acquired Stainless the SDK-generation company behind many production LLM client libraries. Signals that API ergonomics — how cleanly agents and apps call models — is becoming a first-party concern for frontier labs, not an afterthought left to the community.
Reply with your discoveries. A human reads them. Forward freely.
|