Loading today’s stories
Build1 distinct publisher2 min readPublished
Salesforce's Agentforce lets a builder pin an action to a specific response format per connection. The model still decides what to do; configuration decides how it appears. That split is exactly where the 100 percent guarantee applies, and where it stops.
The Engineer · Build desk
leadership
Clara Shih left Meta after her own agents made her stop posting entry-level jobs1 distinct publisher
invest
Salesforce's double digits, minus Informatica: agentic AI is real and still 2% of revenue1 distinct publisher
product
OpenAI starts charging some large accounts only for the jobs its model finishes1 distinct publisher
invest
Salesforce booked more investment gains than operating income last quarter1 distinct publisher
Compiled by The EngineerSomething wrong?How this is made
Start with where the boundary sits. The model keeps everything that needs judgment: reading the conversation, working out what the user wants, deciding which actions to execute [6]. It loses one lookup, and Salesforce places that control in pre- and post-reasoning orchestration [5], which is the only place it can sit if the reasoning step itself is left alone. Abhi Rathna's phrasing is the part worth keeping: Salesforce took the rendering decision out of the LLM's hands, rather than making the model itself more deterministic [7].
The 99 percent in the post is a threshold argument rather than a measured rate [3]. Take it literally anyway, because the shape of the arithmetic explains why sequencing is treated separately. One mandatory render at 99 percent misses one conversation in a hundred [2]. A regulated flow with three mandatory renders succeeds at 0.99 cubed, or 97.03 percent, so roughly three conversations in a hundred go out without something a regulator asked for [1]. That compounding is why the Q&A lists complex sequencing across multi-action and long-running interactions as its own problem instead of a variant of the first one [11].
So what has to be true for the 100 percent to transfer to your build. First, determinism is opt-in. The binding is declared per action and per connection [4], so the guarantee covers exactly the actions someone remembered to enumerate, and an unbound disclosure is still selected by the model [8]. Second, the bound format has to mean something on the channel it lands on. Connections spans web chat, WhatsApp, SMS, phone and kiosks [2], and a rich form has no counterpart on a voice line. What this text supports is deterministic selection of the response format; whether that format survives every channel intact is a separate property the post does not measure.
This is still good engineering, and the reason is the shape of the fix rather than the number attached to it. The failure Salesforce describes is a component that appears in one test interaction and returns plain text in the next [8], which stops a launch rather than degrading it [9]. For regulated customers the standard is that the disclaimer always shows, not that the agent usually shows it [10]. A binding can be read and reviewed; prompt behaviour has to be sampled before you believe it. In my context I would take the enumeration burden over the sampling, and I would budget for paying it again each time a channel is added.
Ranked by verification strength, evidence, and original report placement.
Salesforce's engineering blog published an Engineering Energizers Q&A with Abhi Rathna, Product Management Director for Agentforce, about preserving stochastic LLM reasoning while guaranteeing that specific outputs render correctly 100% of the time.
Rathna's team owns the Agentforce Connections layer, which delivers agent experiences across web chat, WhatsApp, SMS, phone, kiosks and other channels, and covers agent behavior customization, response customization, multi-channel rendering, and structured JSON responses for headless scenarios.
For critical forms, components and regulated disclosures, the post states that 99% reliability was still failure: a required form had to render and a compliance disclosure had to appear.
The render: directive lets a builder associate an action with a specific response format for a particular connection; when that action runs, the Connections layer uses the configured response format instead of asking the LLM to select the UX component.
Deterministic control was added as part of pre- and post-reasoning orchestration to provide more fine grained control on UX rendering.
The LLM could not be removed from the interaction: the reasoning engine still needs to understand the conversation, determine what the user wanted, and decide which actions should execute.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 3, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One in-house account, specific but unchecked
The mechanism is described precisely enough to disagree with — a named directive, a per-connection binding, a stated place in pre- and post-reasoning orchestration — which is more than most vendor engineering posts risk. It is still a single interview with the feature's own product manager, and nobody outside the team has tried to make a form fail to render. The version in front of us also stops mid-sentence at the question about non-Salesforce and headless surfaces, so the broadest claim arrives with no working behind it.
No deployment signal at all
Not one customer, ship date, release stage, or volume number appears. We can say builders are told they can pin an action to a response format; we cannot say the capability is generally available or that anyone outside Salesforce has used it. The only usage detail in the whole piece is negative — agents that could not launch before the fix.
Absolute headline, bounded guarantee
'100% deterministic rendering' promises a property the post itself narrows four answers later: the model was not made deterministic, one decision was moved out of it. Inside that boundary the claim is plausible — configuration, not inference, picks the component. Outside it, the model still decides whether the pinned action runs, and a disclosure that is guaranteed to render correctly when triggered is not a disclosure guaranteed to appear. The gap is the headline's, not the engineer's.
Product owner, company channel
Interviewee runs product management for the feature; venue is his employer's engineering blog, a channel built to attract engineers and to tell regulated buyers their workloads are safe here. That alignment shows in what is quantified and what is not: the pain is described with a round rhetorical 99%, the outcome with an absolute 100%, and nothing in between is measured. The candor about the boundary suggests an engineer wrote the substance; the framing suggests someone else wrote the title.
Uncontested, uncorroborated
Nothing here is disputed, and nothing here is confirmed — one publisher, one voice, zero replication. We are reasonably sure what Salesforce says it built and where the split falls; we are much less sure how it behaves across multi-action and supervised runs, which the post names as the trickiest part and then documents in a paragraph.