Dreamforce Ships Agent Operations While ROI Numbers Lag, Claude Code "Mods" Show How Unclear Feature Status Can Be, and Virginia Stalls on Chatbot Rules

Three threads this week are about the gap between what is announced and what you can rely on: enterprise agent platforms with thin ROI evidence, a developer feature whose status sources disagree on, and a state that has yet to turn chatbot-safety proposals into law. We could not open the primary pages (Salesforce, Anthropic, Virginia) from our environment, so details come from search summaries and secondary coverage.

1. Dreamforce: Agent Operations Over Headline Models

Dreamforce ran September 15 to 17 in San Francisco. Salesforce's own recap, as summarized in search results, describes AIforce as a layer that brings Salesforce data and workflows into the tools people already use: Claudeforce inside Claude (beta), Slackforce inside Slack (general availability slated for October) and Agentforce Coworker inside Salesforce (generally available). Third-party writeups also mention seven pre-built job-specific agents, a long-horizon runtime for goals that last days or weeks, Koa (described as a CRM-focused reasoning model) and an Enterprise AI Harness for governing and monitoring agents.

Two cautions. One independent analysis says neither Koa nor Claudeforce is generally available: Koa is a pilot with a winter target and Claudeforce has no published price. And a MarketScale headline says Dreamforce went all-in on agents "but ROI numbers are still missing." We did not see the article body, so we can't say what figures it found lacking.

The takeaway for builders is where the vendors are putting effort: monitoring, governance and meeting users inside existing tools rather than another chat window. If you run agents yourself, the unglamorous work is the same. Log every tool call, attach a cost to each task, and track the one metric your business cares about.

type TaskReceipt = { task: string; toolCalls: number; tokens: number; resolved: boolean };
const resolutionRate = (r: TaskReceipt[]) => r.filter(x => x.resolved).length / r.length;
const tokensPerResolved = (r: TaskReceipt[]) =>
  r.reduce((s, x) => s + x.tokens, 0) / Math.max(1, r.filter(x => x.resolved).length);

A simple dashboard of resolution rate against tokens per resolved task would give you the ROI number that coverage says is missing.

2. Claude Code "Mods": When Sources Disagree on Status

A Level Up article says Anthropic recently introduced Claude Code "mods," small TypeScript functions that change how the agent behaves, available in the CLI and desktop app. A separate ClaudeFast page describes "function hooks" as a proposed hook type that has not shipped, quoting an official ClaudeDevs post that calls it something the team is exploring. The descriptions overlap, so they may be the same idea under different names, or not. We could not confirm either from Anthropic's docs.

What is established is the existing hooks system: shell commands run at lifecycle points such as before and after a tool call, and the Agent SDK offers PreToolUse and PostToolUse hooks plus a custom permission callback. The idea of running small typed functions as middleware around tool calls, much like Express middleware, is useful regardless of what any one vendor ships. A sketch of the pattern in your own agent loop:

type Hook = (call: { tool: string; args: unknown }) => Promise<'allow' | 'deny'>;
const hooks: Hook[] = [denyDestructiveShell, requireApprovalForPayments];

async function runTool(call: { tool: string; args: unknown }) {
  for (const h of hooks) if ((await h(call)) === 'deny') return { error: 'blocked' };
  return execute(call);
}

Until a feature's status is clear in official docs, don't build a workflow that depends on it.

3. Virginia Delays Chatbot Safety Rules

A Hoodline report says Virginia is weighing chatbot safety rules and delaying action. Per the search summary, two bills that would have required chatbots to disclose they are not human, enforce crisis protocols and verify users' ages died in committee in March 2026, and a state commission is now pushing principles toward the 2027 session. Virginia has no binding chatbot law yet.

Elsewhere, a search summary reports the Eighth Circuit granted xAI an injunction pending appeal on October 2 against a Minnesota law on so-called nudification tools, after the district court denied a preliminary injunction on September 4. We could not verify the court records, and one legal tracker warns that many of its items trace back to company press releases.

For builders, the pattern is that disclosure, crisis handling and age checks keep appearing in proposals even where none has passed. Making them configurable now is cheap: a clear "you are talking to an AI" line, a tested path to human help when a user shows distress, and an age signal your product can act on.

Conclusion

Enterprise vendors are selling agent operations, a developer extension point has two conflicting status reports, and a state with no chatbot law is still deliberating. In each case, rely on what is documented and measurable: your own ROI metrics, official docs and enacted statutes.

— Maya