|
Hey there 👋
I went through this morning's agent launches expecting a new model. There wasn't one. What every launch worth reading had in common was smaller and stranger: the agent learned from the reader's own data, not the open internet. A startup that mines your best reps' calls. A reporting platform grounded in your own filings. An ops tool that builds agents out of your runbooks.
So here is my bias, for what it is worth. The frontier model stopped being anyone's edge months ago. The edge is the pile of proprietary interaction data you already sit on and mostly ignore, and this week a run of companies shipped products that turn that pile into the training set.
The Big Thing
Encore AI thinks your best rep is the training set.
Encore AI, formerly Insait IO, raised a $30M Series A yesterday, led by Team8, Planven, and The Garage. Set the raise aside for a second. The interesting part is the mechanism, which Encore calls Interaction Mining. It studies your top-performing human agents, pulls out the specific moves that close, and deploys those moves as AI agents across voice, chat, IVR, and live form-fill.
The framing is the tell. Most support bots are built to deflect, to keep you from ever reaching a person. Encore is built to do the opposite: convert leads, close applications, recover balances, and upsell, in regulated sectors like financial services where a wrong sentence is a compliance problem. The agents run autonomously or next to a live team, in any language.
Now the grain of salt. "The only enterprise AI platform built to generate revenue" is Encore's own line, so read it as positioning, not a market survey. And several of the banks and insurers in this round started as customers, which means the people vouching for the results also own a piece of the upside. Believe the mechanism anyway, because it is sound: your best closer's recorded calls are the highest-signal training data you own. The percentages can wait for someone outside the cap table.
You probably cannot buy this the week you read about it, and if you run a ten-person team you do not need to. The pattern is free. Before you write a single agent prompt, pull the transcripts from your three best people and read what they do that the average rep does not. That is your spec.
Ship it? Watch the company, mine your own reps now. The lesson of this whole issue is sitting in your call logs, and it costs nothing to go read them.
TechCrunch · PR Newswire
Tour de Headlines
📊 Workiva put audit trails on its agents.
Workiva shipped three agents plus an intelligence layer called Workiva Knowledge that grounds them in each customer's own filings, instructions, and content. The Tie-Out Agent runs consistency checks across a financial report and flags every discrepancy with an explanation for the variance. The Benchmarking Agent reads peers' 10-K and 10-Q filings, builds custom peer groups, and surfaces the disclosure gaps in your own draft. The Sustainability Disclosure Agent drafts and scores ESG disclosures against the ESRS and ISSB frameworks and hands back a gap assessment. Every step leaves a traceable trail, which in high-stakes reporting is the whole game. Availability is limited to the higher solution tiers. Ship it? Deploy if you are already on Workiva. The audit trail, not the drafting speed, is why this one is safe to run in a regulated close.
Workiva launch (via StockTitan)
🔬 Chip design agents learned to check their own work.
At DAC this week, the three big EDA vendors all shipped autonomous agents on NVIDIA's stack. Synopsys demoed a design-verification agent on its AgentEngineer platform that it says takes RTL verification from weeks to hours, up to 50x faster time-to-validated RTL with a 20% coverage bump. Those are Synopsys figures on Synopsys workflows, so hold them loosely. Cadence added its AuraStack super agent; Siemens added self-verifying agents to Fuse. The number is not the story. The pattern is: these agents verify each decision against a deterministic engine before they move on, instead of generating and hoping. Ship it? Watch if you are not taping out chips. Steal the pattern if you build anything, because an agent anchored to a ground-truth checker beats a confident one that guesses.
Synopsys at DAC 2026
🛡️ Your "share" link is a publish button.
For a few days this week you could type site:claude.ai/share into Google and read strangers' shared Claude conversations and Artifacts: resumes, financial spreadsheets, patient names. The cause was mundane. Shared links are public by design and the pages were missing a noindex tag, so once a URL leaked, crawlers did the rest. Anthropic patched its robots.txt and the results were gone by Tuesday. Do not file this under Anthropic. The same thing hit ChatGPT share links last year, and it will hit the next tool too. Ship it? Fix today. Audit every share and export surface in your own agents, add noindex on anything public, and rotate any secret that ever rode along in a shared thread. Pulling a page from Google does not kill a link someone already bookmarked.
VentureBeat · Axios
|
Sponsor
Measure the human signal in every conversation.
Your agents get scored on every call they handle. RapportScore does the same for your people, reading the communication signals in real sales and support conversations so teams can see what lands and adjust before the next one. Built by the crew behind this newsletter.
See your team's score →
|
Tool of the Day
⚙️ Dynatrace Agent Builder (and a live SRE agent)
What it is for: turn your ops runbooks into agents that fix things, with a human still on the trigger.
Dynatrace pushed its observability platform into autonomous operations this week. One piece is live today for SaaS customers on Platform Services: the Cloud SRE Agent, which coordinates remediation across AWS, Azure, and Google Cloud and folds every action into a single auditable record. The bigger piece for builders, a no-code Agent Builder that turns your own environment-specific workflows into custom agents, lands in August alongside an Autonomous SRE Agent. If you run Dynatrace, the try-this is small: point the Cloud SRE Agent at one well-scoped, low-blast-radius remediation, keep a human approval on the trigger, and read the audit record after each run before you widen its reach. That record is the part that makes autonomy survivable in production. Ship it? Try it on one boring, reversible incident type first. Prove the log and the rollback, then hand it something that matters.
Read the Dynatrace announcement
Worth a Click
- Encore AI, the mechanism. How Interaction Mining turns your best closer's calls into a deployable agent, in Encore's telling. TechCrunch
- Workiva's launch. The three governed agents and the Knowledge layer, with the audit-trail claims spelled out. Workiva / StockTitan
- Autonomous EDA at DAC. A strategic read on what self-verifying design agents mean for high-stakes engineering. Futurum
Step back and the week has a quiet through-line. Last week the story was the governance layer: OpenAI, Google, and Anthropic all selling the wrapper that keeps an agent alive in production. This week the story moved one layer down, to the data you feed it. Encore mines your reps. Workiva grounds its agents in your filings. Dynatrace builds them out of your runbooks. Even the chip vendors are anchoring agents to their own verification engines.
If I had to bet, this is the more durable moat of the two. A competitor can rent the same model you rent and buy the same governance tools you buy. What they cannot get is your interaction exhaust: the call recordings, the closed tickets, the filings, the runbooks you have been generating for years and mostly throwing away. Start treating that exhaust like the asset it is. Label it, keep it, and feed it to the agent before you reach for a bigger model.
Ron
|
