The latency problem that rewired everything
Voice agents couldn't wait for frontier model responses. That constraint forced Decagon's founders, Ashwin Sreenivas and Jesse Zhang, to abandon the OpenAI and Anthropic APIs they'd launched on in 2023. By 2025 the product worked; enterprise customers including Duolingo, Oura, and ClassPass were resolving 70 percent of chats, as Decagon's website reports, and cutting costs 95 percent, Decagon's case studies show, but voice introduced a hard ceiling. "The only way to get latency down, but also kind of make our agent operate the way we want it to, is to use smaller models," Zhang said in a July 2026 a16z podcast. The team began fine-tuning open-source models for specific tasks. Today 90 percent of inference runs on those models, supported by an internal research group (Decagon Labs) that functions as a model factory for the customer-service use case.
That shift rewired daily engineering work. The research team exists because "we need people that are really good at taking these open source models and tuning them," Zhang said. Training infrastructure and evaluation pipelines are built in-house; labeled data and dataset-diversity measurement are bought. The team constantly trains net-new models and deprecates old ones as the frontier advances. Frontier models still handle auxiliary tasks, such as autopilot workflows requiring broad reasoning, but the core conversation loop runs on smaller, task-specific models that are cheaper, faster, and more controllable.
Engineering without the forward-deployed trap
Decagon's forward-deployed engineers operate under the same bar, reporting structure, and often the same teams as product engineers. The divergence that exists at Palantir or Databricks, where forward-deployed and product roles eventually split into separate ladders, hasn't happened here. "Every enterprise customer request is a potential product feature that must be built to scale across the entire customer base," said Sunny Rekhi, CTO of Forward Deployed Engineering. The scarce skill isn't writing code fast; it's exercising restraint. When a Fortune-20 customer demands a feature, the temptation is to prompt an AI coding agent and ship a one-off. Decagon's engineers are measured on whether they architect the solution so the next customer inherits it.
This philosophy drove a role split in 2025. The original "agent software engineer" title divided into two lanes: agent builders, who configure agents through a UI using Agent Operating Procedures (natural language instructions that compile into structured logic), and agent software engineers, who productize custom requests into platform capabilities. The split let the company scale from 50 to 500 employees in roughly one year, Sunny Rekhi reported, without sacrificing speed or quality. Agent engineers sit at the center of product strategy, with meaningful ownership over product direction, platform design, and customer outcomes. Their work transitions from concept to production in days, directly impacting deals and revenue.
Scoping discipline as tempo
Decagon automates the highest-ROI intent first so value appears in weeks, not months. Engineers act as advisers, not order-takers, using cross-customer pattern recognition to decide what belongs in the platform. "Anything that has to be done manually by engineering should be upstreamed into the product," the company's operating principles state. Custom becomes self-serve. Vertical domain knowledge compounds. The engineering setup (connecting backend systems, configuring APIs, building safeguards) happens once, then CX teams take operational ownership through the AOP layer. The next deployment is faster than the last one.
Values that show up in the architecture
Decagon's stated values read like a reaction against the consulting-style deployments that dominate enterprise AI. "We value decisive action and speed over prolonged deliberation and planning," the company declares on its site. The principle shows up in how the team talks about iteration: "Every update to your agent shouldn't require an engineering sprint or vendor ticket." That line isn't aspirational — it describes the product architecture. Agent Operating Procedures (AOPs) let customers rewrite workflows in natural language, and the platform ships those changes without a code deploy. A customer that migrated from Sierra told the a16z podcast that Sierra's forward-deployed-engineer model capped them at three new journeys per year; Decagon's "glass box" approach let the same team launch seven in a month.
The glass-box metaphor is the clearest operating principle the founders repeat. On the July 2026 a16z podcast, CEO Jesse Zhang contrasted Decagon with competitors who treat the agent as a black box: "We like to call this a glass box approach instead of a black box." The distinction is structural. Sierra's FDEs own the configuration; Decagon's product exposes the logic so the customer's own team can audit, test, and extend it. That same principle shapes the internal org. Deployed engineers at Decagon don't build one-offs: they "go in and understand as we're working with an enterprise, what are the things that they need that the product does not do today, but the output of that is not a one-off thing that is just built for that customer. It is something that is contributed to core product." Agent product managers run the same loop: they sit with customers to learn where the product breaks, then feed those gaps back into the roadmap.
Model pragmatism over lab loyalty
Speed as a value appears in the model strategy, not just the product cycle. When Decagon started, the goal was "just to get something working," so they used frontier APIs. At scale, latency and cost forced a pivot: "about a year ago, we were like, okay, let's start using these open source models." Decagon Labs functions as an internal model factory: "compress the time between new model coming out and you know useful fine-tuned to our task model kind of popping out the other end." The team fine-tunes for the customer-service use case, not per customer. "Most of the fine-tuning we do is like customizing it for our use case," Zhang said. "It's worth it for us to do it because that's all we do."
That pragmatic streak extends to how the founders talk about the application layer's future. Sreenivas rejects the narrative that foundation-model labs will absorb every vertical: "Even once you have AGI, agents are going to need somewhere to store work and pull information from and reason about things. I don't think software as a whole in any meaningful way is going away." The bet is that business logic — rebooking three passengers on a canceled flight, running compliance checks, orchestrating a proactive outreach campaign — lives in the platform, not the model. "How the business logic gets captured by this AI... has nothing to do with the models themselves."
Culture coheres around velocity, not hours
Internally, the culture coheres around product velocity, not performative hours. "Working hard is an effect of wanting to build a good product, not a goal," Zhang said on the podcast. "No one is mandated to be in on weekends; people are in the office 'just so that we can maximize communication.'" The phrase he used, "It's a we're-all-in-this-together to get this across the line," matches the customer testimonial on Decagon's site: "The Decagon team consistently finds ways to make things work. They really understand what we're asking for, get creative in how we solve problems, and keep pace with how quickly we want to stand things up."
The founders also treat external communication as a forcing function for clarity. Zhang described X as "a single timeline that everyone reads" that "kind of mind-controls everyone into thinking about the same thing." His rule: post industry theses, not company promotion; use AI for brainstorming topics, not writing; never cross-post LinkedIn and X because "very few things do well on both." The discipline mirrors the product philosophy — strip the wrapper, expose the reasoning, ship the substance.
The operating system is coherent: glass-box transparency, productized delivery over services revenue, model pragmatism over lab loyalty, and a communication cadence that aligns the team to the same problems. That system filters for the people who survive the interview loop.
Inside the interview loop
Glassdoor shows two distinct Decagon interview datasets: one with 38 questions and 42 anonymous candidate reviews, another with 6 questions and 8 reviews. The volume alone signals a company hiring steadily enough to generate a pattern — not a handful of one-off roles. The technical loop centers on real Decagon problems, such as agent orchestration, latency budgets, and evaluation frameworks, rather than abstract textbook scenarios, reflecting the daily constraints the team faces.
Decagon's own product metrics (70% chat and voice resolution, 80% deflection rate, Duolingo's case study on Decagon's site found, 95% cost reduction for some customers) imply an engineering culture that measures ruthlessly and iterates publicly. Candidates who treat the loop as a collaboration, ask clarifying questions mid-problem, and propose trade-offs unprompted tend to advance. That behavior maps directly to how the team operates: ship, measure, improve, repeat.
What the pay data shows
| Role | Location | Salary Band |
|---|---|---|
| Engineering Manager, Agent Orchestration | San Francisco | $280k–$430k |
| Director of Sales, Enterprise | San Francisco | $336k–$420k |
| Engineering Lead, Strategic Deals | New York / San Francisco | $200k–$400k |
| Engineering Manager, Agents | San Francisco | $200k–$400k |
| Research Engineer | San Francisco | $200k–$400k |
Decagon's compensation structure sits at the upper end of the AI startup market. The clearest picture comes from the Zero G Talent board's first-party data: across 107 salaried roles, the band runs $141k–$400k with a median of $280k. That aligns with the H-1B disclosure for FY 2026, where the median wage for Decagon AI Inc. landed at $263,500 and the 75th percentile at $320,000. Third-party aggregators report wider spreads; levels.fyi cites a $510k–$759.5k total-compensation span for software-engineering titles, but those figures blend base, bonus, and equity into single numbers and include non-U.S. locations, making them less useful for benchmarking the core U.S. engineering cohort.
Equity grants follow a standard Silicon Valley pattern: incentive stock options (ISOs) and non-qualified stock options (NSOs), per OptionsAhoy's plan analysis. The same source notes the critical operational detail: a 90-day post-termination exercise window for vested ISOs, after which unexercised options are forfeited. Vesting schedules and refresh-grant cadence are not disclosed in any public filing or company communication; the only grounded statement is that the 90-day window is the mechanism that converts paper equity into exercisable rights upon departure.
Benefits are spelled out on Decagon's own careers page: medical, dental, and vision coverage with 100 percent dependent coverage paid by the company. Built In's benefits summary corroborates the health-insurance core and adds retirement-plan access and a vacation policy, though the latter's specifics (accrual rate, carryover, minimum days) are not published. Glassdoor's 11 salary reports across eight roles are too sparse to derive reliable percentile bands, but they do not contradict the board's $141k–$400k corridor.
Taken together, the data portrays a compensation philosophy that leads with high base salaries, clustered around the $280k median, layered with conventional ISO/NSO grants and a benefits package that removes dependent-care cost from the employee's ledger. The missing pieces (vesting schedule, refresh policy, bonus targets) are the same ones most early-stage AI companies keep off the public record; candidates should ask for them explicitly in the offer stage.
That gap is not noise — it is the company's operating contract written in numbers.
Who thrives here and who doesn't
The Glassdoor data tells a clear story: 257 reviews, 3.9 overall, 75 percent would recommend the company. But the sub-scores reveal the fault lines. Culture & Values, Work-Life Balance, Senior Management, and Career Opportunities all sit between 3.7 and 3.8. Compensation sits at 3.2. That gap is not noise. It is the company's operating contract written in numbers.
Decagon is a bet on equity upside and career acceleration, not a balanced or lifestyle company. The research makes this explicit. Founders Jesse Zhang and Ashwin Sreenivas (both ex-Google, both builders of Helia, acquired by Scale AI) still run technical reviews. Engineers deploy to F500 customers multiple times per week. Forward-deployed engineers travel to enterprise sites and ship fixes within days. The pace is described by multiple reviewers as "the fastest I've ever worked," framed positively by most, cautioned by some.
Five profiles that thrive
Builders who want applied AI impact. Decagon's agents handle live support traffic across 100-plus enterprise customers including Avis Budget Group, Block, and Deutsche Telekom. Shipped code reaches end-users within hours, not quarters. Engineers see the effects of their work in production within days. If you want to work on AI that real users depend on — not research for its own sake — this is one of the cleanest examples in the market.
People who optimize for equity and career acceleration. The company is explicit that base comp is not top-of-market. Frontier labs such as Anthropic and OpenAI pay more cash. Decagon's case is built around equity upside at a $4.5 billion valuation (up from $1.5 billion at Series C in mid-2025) plus the career-acceleration value of owning large surface area early at an AI agents company that is clearly winning. Early employees who joined pre-Series C have already experienced significant paper gains as the valuation rose from roughly $65 million at seed in 2023 to $4.5 billion in January 2026.
Engineers who want to ship fast. Deploy cadence is fast. Feedback loops from forward-deployed engineers and F500 customers are short. If you've been frustrated by approval gates and multi-week planning cycles elsewhere, this is a breath of fresh air. The tech stack (TypeScript, Python, React, LLM tooling) is modern but pragmatic, chosen for velocity.
Forward-deployed engineers and generalists. Decagon invests heavily in FDEs: engineers who sit with customers, translate enterprise requirements into agent configurations, and ship backend changes to match. If you like the mix of customer-facing work and real engineering, few places do this as intensely. The role is unusually impactful: you ship code and watch it go live in an F500 customer's support workflow the same week.
People comfortable with SF-centric, in-person culture. Decagon is not a remote-first company. Most engineering, product, and GTM roles are on-site or hybrid with strong preference for in-person collaboration. Forward-deployed roles involve travel to enterprise customer sites but still route through San Francisco. Fully remote roles are rare. If you thrive in office environments with high density of talent, this is a plus. If you don't, it's a dealbreaker.
Who struggles
People who prioritize work-life balance above all else. The 3.7 WLB score reflects real trade-offs. Work hours trend long. The company is explicit: this is a Series D sprint to category leadership, not a lifestyle job. Long hours are a consistent theme in reviews, in line with Decagon's positioning as a fast-paced SF AI startup.
People looking for a remote-first company. The SF-centric model is not negotiable. With headcount roughly tripling in 18 months (from roughly 100 to 300-plus), process and formal structure are still catching up to scale. Remote work would add friction to an already high-velocity, in-person collaboration model.
People whose financial situation requires top-of-market cash compensation. The 3.2 comp score is the single biggest gripe in reviews. Base salaries for senior engineers in San Francisco are generally in the competitive-but-not-market-leading range. Benefits are standard-for-SF: health coverage, equity with typical four-year vesting, team lunches. This is not a company known for perks theater — the culture is unapologetically work-focused.
The mismatch is not subtle. Decagon's board salary band runs $141k–$400k (median $280k) across 107 salaried roles, competitive for a Series D startup but trailing frontier labs. Recent postings show Engineering Manager, Agent Orchestration at $280k–$430k, Zero G Talent's board reports; Director of Sales, Enterprise at $336k–$420k; Research Engineer at $200k–$400k. Equity carries most of the upside. If you need the cash now, the math doesn't work. If you can bet on the next leg of growth — the company is hiring aggressively across engineering, forward-deployed engineering, product, design, and GTM following the $250M Series D — the trade-off may be the best career move you make.
For flat-organization lovers at a similar stage, Linear and Granola offer alternative cultural models. If the trade-offs are dealbreakers, consider Linear (4.4 WLB), HubSpot (4.1 WLB), or the frontier labs Anthropic and OpenAI for higher cash comp. Decagon is not for everyone, and the culture is fairly upfront about it.
Working in frontier tech? Zero G Talent tracks the openings: see every open Decagon role, browse frontier tech jobs, the companies hiring, and the people building the field.