Product Roadmap · July 2026

Finishing what the foundation made possible

Building on June's sandbox runtime, July finishes the carryover: grounding the migrated agents in our own knowledge, making the chat assistant trustworthy enough to launch, and metering every action with per-action credit tracking.

Themes 3P0 0P1 9Source June carryover

Next · library + assistant→Later · credit tracking

Overview

next · later

July spends what June built. With the agents on an isolated, traced sandbox runtime, the three themes that ride on that foundation can land - grounding the agents in the knowledge base, launching the chat assistant, and metering every action with per-action credit tracking that reconciles into the Stripe billing system shipped in June. All three carry over from the June roadmap; this is a continuation month.

Themes

all carried from June

P0 priorities

continuation month

P1 priorities

library · chat · credits

Carried from June

ride the runtime

Next · P1

Theme 1Asset library
Theme 2AI chat assistant

Later · P1

Theme 3Credit tracking

Asset library

P1P1P1

The asset library today is a passive store of files - agents cannot draw on what is in it, so their output stays generic instead of grounded in our own brand voice, templates, and domain knowledge. Turning it into a real knowledge base means giving it retrieval infrastructure and wiring it into each agent as grounding context, so every agent works from the same curated source of truth. It builds directly on June’s migrated agents.

Build the knowledge-base infrastructure

Turn the asset library into a retrievable knowledge base: ingestion, indexing and embeddings, and a retrieval API the agents can query.

Wire the knowledge base into the migrated agents

Have the rebuilt agents ground on the library at run time.

Knowledge base across all agents

Generalise grounding to every migrated agent, with the management UX to keep it current.

Theme 1: turning the asset library into a retrievable knowledge base through ingestion, embeddings and a retrieval API that grounds every agent — Turning the asset library into a retrievable knowledge base through ingestion, embeddings and a retrieval API that grounds every agent.

AI chat assistant

P1P1P1P1

The chat assistant has been in build for a while - it is the surface that lets users drive the product in natural language, including the bulk operations that are painful to click through today. What is left is what makes it trustworthy enough to launch: run it on the same Claude agent stack as everything else, let it act through the ac CLI, keep it inside guardrails, and hold it to an eval bar we trust.

Migrate the assistant to Claude agents

Move it onto the Claude agent stack and the sandbox runtime, consistent with the rest of the platform.

Give the assistant ac-cli access

Let it act across the product through the ac CLI surface, which is what makes bulk operations possible.

Add guardrails

Bound what the assistant can do and say: permission scoping, confirmation on destructive actions, and clear limits.

Improve assistant evals

Expand evals over the common flows so we can trust output quality before opening it up.

Theme 2: the chat assistant on the Claude agent stack acting through the ac CLI, bounded by guardrails and held to an eval bar — The chat assistant on the Claude agent stack acting through the ac CLI, bounded by guardrails and held to an eval bar.

Credit tracking

later

P1P1

Accurate per-action credit accounting is what makes usage-based pricing precise. June’s runtime migration is what makes it tractable: the sandbox runtime’s telemetry is the natural seam to meter every provider and LLM call. Today only the extraction LLM calls are logged; Exa, Parallel, Firecrawl, and Hunter calls are untracked. The June Stripe billing system shipped on Stripe’s own usage records; this work makes the numbers precise and reconciles back into it.

Instrument per-action credit tracking

Wrap each provider call (Exa, Parallel, Firecrawl, Hunter) and LLM call (strategist, extraction, qualifier) with per-action accounting as the agents move onto the new runtime.

Generalise tracking, then reconcile with billing

Lift the accounting into a shared mechanism every migrated agent emits, then reconcile the proven per-action numbers back into the June Stripe billing balance. Instrument first, refine billing later.

Theme 3: wrapping each provider and LLM call with per-action accounting, lifted into a shared mechanism that feeds credits-based billing — Wrapping each provider and LLM call with per-action accounting, lifted into a shared mechanism that feeds credits-based billing.

KEYThe through-line

June moved the agents onto a runtime we can isolate and trace; July spends that. The knowledge base grounds them, the assistant puts them in users’ hands, and credit tracking meters every action - reconciling into the Stripe billing system that shipped in June. The foundation pays off.