Three launches from the past week caught my attention, partly for the products themselves and more for what they sketch together. None of them is a category-defining release on its own. Read alongside each other, though, they map out something I've been waiting to see for a while: what an application-layer product looks like once agents are real members of the workflow rather than external tools bolted onto it.
As the infrastructure layer for agents matures, the more interesting question is what the layer above it looks like. The application-layer products of conventional software — issue trackers, code review, dashboards, project management — were designed for humans collaborating with each other. If agents are first-class collaborators rather than external tools, the shape of those products has to change. The three releases below point at what that change looks like in practice.
A note on what follows. Each item below is a recent release, several within the last few days. Mentioning them is not a recommendation. None has a long community track record yet; treat this as a watchlist, not a buy list.
Batta: shift the security review left, to the agent's plan phase
Batta is an open-source tool that intercepts an AI coding agent at its plan phase — the moment when the agent has decided what it intends to do but hasn't yet written any code — and runs an automated security review against the plan itself. The project is at https://github.com/Dolevco/batta-ai.
The standard place to put security review for AI-generated code is after the diff exists: a linter, a SAST tool, a human checking the patch before it merges. Putting the review at the plan phase moves failure detection upstream of the expensive work. If the agent's plan is to disable a security check, exfiltrate a credential to logs, or implement an authentication bypass, you want to catch that intent before two hundred lines of code reify it. The bigger lesson, beyond Batta itself, is that the right place to gate an agent's behavior is upstream of the artifact, not downstream of it. For vertical LLM SaaS in compliance-heavy domains — healthcare, finance, defense, anything with auditable workflow — this is the kind of constraint your customers' regulators will eventually expect you to encode in the workflow rather than catch in post-deployment review.
Paca: a project board built around the assumption that agents are scrum members
Paca launched as a lightweight, open-source project-management platform that treats AI agents as scrum members rather than external bots. Agents in Paca get assigned to sprints, pick up backlog tickets, contribute to BDD specifications, and update system design documents alongside human teammates. The platform is built on Model Context Protocol (MCP), Claude Code skills, and the OpenHands SDK. The project is at https://github.com/Paca-AI/paca.
The gap Paca is addressing is real and underacknowledged. Every team that has actually shipped a feature using an agent has hit the same problem with their tracking system, whether it's Jira, Linear, or something custom: the agent is modeled as either an external "bot user" or not modeled at all. The result is that agent work is invisible to the team's planning cadence, or visible only as awkward bolt-on comments on tickets a human owns. Building the planning surface around the assumption that agents are members of the team — they have a velocity, they pick up work, they participate in review — changes how the team plans, retrospects, and assigns. Whether Paca specifically becomes the standard or just an early example of the shape doesn't matter; the shape itself is what to watch. For any vertical SaaS that includes a workflow or project-management surface, the corresponding question is whether your product treats agents as participants or as integrations.
BitBoard: analytics dashboards as durable, agent-generated artifacts
BitBoard (a YC P25 company) launched an analytics workspace where AI agents are first-class authors. They can push data, generate analyses, and build dashboards that persist as durable, traceable assets instead of disappearing into ephemeral chat threads. The product is at https://bitboard.work/.
The unsung problem with chat-driven analytics is that the analysis disappears the moment the conversation ends. Every analyst has had the experience of asking a model to compute something useful, getting a clean answer, and then having to ask again next week because the artifact didn't survive the session. Making the agent's output land as a saved, versioned, shareable dashboard turns the model from a clever ephemeral tool into something the team can build on. For vertical SaaS with any analytics surface, the comparison customers will silently make is to spreadsheets: if a customer can ask the embedded agent for an analysis but can't save the analysis as a dashboard the team can rerun next quarter, you're shipping a regression relative to Excel.
What this layer is actually about
Each of these is a small bet on the same question: what does this application category look like once agents are real participants in the workflow rather than external tools bolted onto it? The interesting move for any vertical LLM SaaS team is to ask the same question of their own product. If a customer's workflow now includes an agent doing some fraction of the cognitive work, where in your product is the agent a first-class participant — visible in the planning, accountable in the audit log, modeled in the data — and where is it stuck pretending to be an external API integration? The gap is where the next round of product work probably lives.
About Heavy Chain
Heavy Chain works with engineering organizations to deploy AI-native PDLC and SDLC — the Etc framework, together with the tooling, gates, and review rituals that make it real — into existing teams. The harder half of the work is change management: helping the new lifecycle stick without breaking what already works. If any of the threads above touch your business, that's where we come in.