Signal Watch: Six Agentic Roles, Three Different Control Layers

Six companies posted agentic design roles this cycle, and all six postings use the word "agentic." The word tells you less than the control layer does: which part of the agent's behavior the designer actually owns.

Three layers show up across these postings:

Generation infrastructure. The designer shapes what the AI can see and produce: structured content, machine-readable design guidance, the data pipelines that feed generation. It matters, but it sits below the interaction layer, which is where your proof is strongest.

Domain control policy. The designer decides what the agent automates, what it explains, and what it escalates. This is the trust boundary, where the human stays in the loop and under what conditions. Your TinyFish platform work and "Trust Is the New Interface" map here directly.

Agent orchestration. The designer builds the surfaces humans use to govern multiple concurrent agents, recover from failures, and keep fleet-level visibility. Your TinyFish agentic platform experience is relevant proof here. Databricks Omnigent is the cleanest example of this problem in the current cycle.

Amplitude — Principal Product Designer, Wave

Trigger: Principal Product Designer posting for Wave, Amplitude's proactive product agent. Surfaced ~Sep 29. High confidence.

Window: Wave is in closed beta and rolls to customers in H2 2026, so this is pre-GA, 0-to-1 definition work. The posting is ~12 days old. Hot, no decay yet.

Snapshot: Amplitude is a public analytics company ($200M+ ARR range) betting its next growth phase on Wave, an agentic system that monitors behavioral data, identifies product opportunities, reads the codebase, orchestrates coding agents to ship changes, runs experiments, and measures outcomes. Whoever defines Wave's interaction model is defining how product teams move from doing analysis themselves to governing an agent that does it for them.

What this trigger reveals: Closed beta means the core interaction model is still being shaped. Amplitude's June 2026 engineering post frames the central design problem as "the shift from controlling to steering": users approve, reject, and redirect instead of executing. The posting says the designer "may ship front-end code directly," which fits Amplitude's code-forward design culture. You'd be defining a product, with no shipped surface to iterate on yet.

Challenge taxonomy: 0-to-1 product definition at the orchestration layer. The designer governs a multi-step loop (behavioral signal, opportunity identification, code generation, experiment, measurement) in which the human's job is oversight and course correction across concurrent agent actions. Your TinyFish 0-to-1 agentic platform build is direct proof. The trust essay's progressive delegation framework applies to the approval surfaces.

Landscape context: Horizontal AI hasn't covered analytics companies pivoting to agentic platforms yet. That gap is worth filling, because Mixpanel and Heap face the same strategic question. Wave's differentiator, the closed loop from behavioral insight through coding agent to measured experiment, also hasn't been benchmarked against coding-agent tools like Cursor or Windsurf, which have no analytics layer at all.

Rubric:

Verdict: Act. Company 10, Role 14. Pre-GA agentic product, 0-to-1 timing, code-forward culture: that lines up with your TinyFish proof. Governing a multi-step autonomous loop is the orchestration-layer problem where your experience is most differentiated, and the problem is new enough to produce a strong portfolio case whatever Amplitude's stock does.

Warning signs: Amplitude's stock has been volatile. Wave's pitch, a proactive agent that reads your codebase and ships changes, will run into serious enterprise trust resistance. And the emphasis on shipping front-end code means the technical execution bar is high and specific.

Databricks — Staff Product Designer, Omnigent

Trigger: Staff Product Designer posting for Omnigent, Databricks' multi-agent governance harness. Surfaced ~Sep 29. High confidence.

Window: Omnigent entered managed beta June 17, 2026. The posting is ~12 days old. Hot.

What Omnigent actually is: A meta-harness. It sits above existing coding agents (Claude Code, OpenAI Codex, Inflection Pi, custom agents) and wraps them in a shared governance and orchestration layer. Users keep the agents they already use and add Omnigent on top. Governance controls (filesystem restrictions, network policies, cost limits, human-in-the-loop gates) are enforced structurally at the orchestration layer rather than through prompt compliance.

The posting names the interaction surfaces: co-driving, forking, approvals, session sharing, live status, error recovery. Taken together, they describe how a team navigates and governs a concurrent fleet of agents. Think air traffic control more than pair programming.

Challenge taxonomy: Agent orchestration in its purest form. The designer defines how engineering teams supervise, govern, and recover from failures across multiple concurrent autonomous coding agents. Your TinyFish agentic platform work is relevant, though Omnigent's problem is a fleet, which goes beyond single-agent orchestration.

Landscape context: The Design Roles Board already covered Databricks with a full rubric and engagement package. What this adds is the control-layer placement that the earlier evaluation didn't include, specifically the meta-harness architecture that separates Omnigent from the single-agent coding tools covered in Horizontal AI.

Rubric:

Verdict: Act. Company 12, Role 14. Highest company score this cycle, and fleet governance is the most architecturally distinct orchestration-layer role in the current scan. Because Omnigent is open-source, the work would be publicly referenceable, which matters for portfolio. The Design Roles Board engagement package applies.

EvolutionIQ — Staff Product Designer, Agentic Platform

Trigger: Staff Product Designer posting surfaced Oct 9. High confidence.

Window: 0-2 days old, peak freshness. Medhub for Casualty went GA in Q3 2025 and more products are planned. The design system for how EvolutionIQ's AI interacts with CCC's much larger adjuster base hasn't been settled.

Snapshot: EvolutionIQ built AI-powered claims guidance for disability and workers' comp insurers (9 of the top 15 disability carriers). CCC Intelligent Solutions acquired it for ~$674M in January 2025. CCC is public (CCCS) and serves 300+ auto casualty insurers. The acquisition thesis is to extend EvolutionIQ's AI synthesis and next-best-action guidance from disability into CCC's auto casualty network. The first integration milestone has shipped, with more in the pipeline.

The core interaction: the AI synthesizes large sets of medical records with citations to primary sources, then recommends next-best-action steps to claims adjusters. The adjuster reviews, applies judgment, and decides.

What this trigger reveals: This designer works across two audiences: EvolutionIQ's existing disability and workers' comp customers, and CCC's auto casualty network, where integration products are actively shipping. On the Q3 2025 earnings call, CCC called the auto casualty integration "a central pillar of our investment thesis." That makes this a cross-platform integration charter arriving during active expansion.

Challenge taxonomy: Domain control policy in a regulated vertical. The designer decides what the AI synthesizes, what it recommends, and what it escalates to human judgment, inside a claims workflow where errors carry financial and legal consequences. Your CarrierIQ work (inspect-then-authorize in insurance) is structurally parallel proof. The trust essay's return conditions, the points where automated authority reverts to human control, apply directly to claims escalation design.

Landscape context: Vertical AI hasn't covered insurtech claims automation yet. CCC's $674M acquisition and the active cross-platform integration make a strong candidate for that coverage, because the design problem (AI synthesis with citation provenance for regulated decision-making) keeps appearing across healthcare, legal, and insurance. Your CarrierIQ citation-chain work is the same structural problem in a different domain.

Rubric:

Verdict: Act. Company 9, Role 13. The domain-control-policy problem matches your CarrierIQ and trust-essay proof directly. And the integration moment, EvolutionIQ's AI expanding into CCC's 300+ insurer network, is the kind of charter that produces strong portfolio cases.

Warning signs: RSUs in CCCS give you stability but limited upside compared to a pre-IPO grant, and the posting doesn't specify whether the RSUs are CCCS shares or something else. Integration politics between an acquired startup and a large enterprise parent are real. How high design influence can go depends on how much autonomy the EvolutionIQ unit keeps. Ask about reporting structure and decision authority in the first conversation.

Félix — Staff Product Designer

Trigger: Staff Product Designer posting, late September (the ATS doesn't expose a publication timestamp). Moderate confidence on timing.

Window: Two signals, at different points in their decay. The $200M Series C closed September 1 ($87M equity led by a16z, $113M credit facility from General Catalyst). That's day 40, well past the funding signal's hot window and into Ignore territory on its own. The job posting is the live signal: ~2-3 weeks old, sitting near the hot/Watch boundary at roughly day 14-21. Read the window off the posting, not the funding.

Snapshot: Félix is a WhatsApp-native money transfer platform for US Latino immigrants sending remittances to Latin America. Users send money by texting through WhatsApp; there's no separate app. It runs on blockchain stablecoins with AI-driven verification across 11 Latin American markets, with a reported $8B+ in cumulative transfers. The Series C funds expansion from remittances into lending, savings, and budgeting through a "Cognitive Financial Companion," a conversational AI assistant embedded in WhatsApp.

A new 1% federal excise tax on cash-funded international transfers exempts digital-rail platforms like Félix, which hands it a structural cost advantage over Western Union and cash-agent competitors.

What this trigger reveals: The capital is earmarked for moving from a single-product remittance platform to multi-product financial services. A designer arriving now defines the interaction model for lending, savings, and budgeting inside a conversational WhatsApp interface, for a population whose trust in financial institutions may be low and whose primary digital channel is the one Félix already occupies. The compliance design load (TILA, FCRA, and AML disclosures inside a chat interface) is substantial.

Challenge taxonomy: Trust architecture in a regulated vertical, delivered through a conversational channel. The designer defines compliance disclosures, consent flows, and recovery patterns for irreversible financial transactions within WhatsApp's constraints. Your Alle work (regulated healthcare transactions, patient consent, provider compliance in a consumer app) is adjacent proof. The trust-design problem is structurally similar, but the financial regulatory regime, the population, and the conversational channel are all different.

Control layer: Domain control policy. The designer decides what the AI financial companion automates (routine transfers), what it explains (credit terms, fee structures), and what it escalates (unusual transactions, compliance flags), all in text, without the visual affordances of a native app.

Landscape context: Vertical AI hasn't covered fintech for underserved populations, and AI Capital hasn't published an analysis of a16z's pattern in conversational AI platforms. Both are gaps. The stablecoin-rail regulatory advantage (the excise tax exemption) is the kind of structural tailwind that would sharpen a Vertical AI evaluation of the remittance corridor.

Rubric:

Verdict: Act. Company 10, Role 13. The regulated-vertical trust problem matches your Alle proof and the trust-essay framework. Designing compliance and consent inside WhatsApp is a hard constraint, and hard constraints produce differentiated portfolio cases. The a16z backing and the regulatory tailwind support the trajectory score.

Warning signs: Team size data is contradictory (sources range from 79 to 442 employees), and design headcount is completely unknown from public data. Total funding figures are reported inconsistently; the discrepancy most likely reflects whether the $113M credit facility gets counted. Ask about design team size, reporting structure, and the actual scope of the "Cognitive Financial Companion" in the first conversation. The distance between the vision (full financial services AI) and the current product (remittances and top-ups) tells you whether this is 0-to-1 definition or incremental feature work.

Anthropic — Three-Role Cluster (Design Systems, Safeguards, Developer Platform)

Trigger: Three Product Designer postings between Sep 23-28: Design Systems, Safeguards, and Claude Developer Platform. High confidence.

Window: 13-18 days old. Still hot. Joel Lewenstein (Head of Design) has said publicly that he is "doubling the product design team" and that "every team that I have designers on is understaffed." Read this as org expansion rather than isolated backfills.

Snapshot: Anthropic builds Claude. Lewenstein leads the design org, with Jenny Wen (Head of Design, Claude) and Meaghan Choi (Head of Design, Claude Code and Claude Cowork) as the publicly identified sub-leaders. Teams work in pods of 3-5 where engineers make design calls and designers push code. Quality gets evaluated in live, working code.

What this cluster reveals: Three roles in five days signals an org buildout. But the three roles sit at different control layers, so they're three different opportunities with different proof requirements. Evaluate them separately.

Safeguards sits at the domain control policy layer. The designer defines what Claude can do, what it explains, and what it refuses or escalates: trust architecture at the frontier, governing how a general-purpose AI interacts with hundreds of millions of users. Your trust essay's framework (progressive delegation, return conditions, mandate maintenance) applies directly, and your Alle and CarrierIQ proof (regulated trust boundaries) transfers. Strongest match of the three.

Developer Platform sits at the agent orchestration layer. The designer defines how developers build with Claude's APIs, manage agent deployments, and monitor agent behavior. It falls under Choi's mandate (Claude Code, Claude Cowork). Your TinyFish agentic platform work is relevant. Moderate match.

Design Systems sits at the generation infrastructure layer. The designer makes Anthropic's design system legible to Claude itself through machine-readable guidance, metadata, and evaluations. The posting mentions "shipping package changes" and "lint or automated UI-quality checks." It's a novel problem, but it lives at the infrastructure layer, where your proof is thinnest. Weakest match.

Landscape context: Horizontal AI has covered Anthropic's position in the frontier AI landscape. What sharpens this evaluation is that Safeguards is a trust-architecture mandate at the company most publicly identified with AI safety, and Anthropic's safety brand as a product differentiator shapes how the portfolio case would read. AI Capital hasn't analyzed Anthropic's latest funding structure; the equity upside question below would benefit from that coverage.

Rubric (company-level, shared):

Role-level (Safeguards — strongest match):

Verdict: Act on Safeguards. Company 11, Role 14. Domain control policy is your strongest proof area, and doing that work at Anthropic's scale would give you a portfolio case nobody else in your competitive set has.

Watch on Developer Platform. Role score ~12. Your TinyFish proof applies, but you'd be competing with many candidates who have developer-tools backgrounds.

Watch on Design Systems. Role score ~11. The problem is novel, but it sits at the generation-infrastructure layer where your proof is thinnest, and the coding bar (shipping package changes, lint checks) is high and specific.

Warning signs: The external title "Product Designer" doesn't confirm internal level, which affects both comp and scope expectations. The pod model means new designers have to establish value fast. Wen has described the traditional design process (discovery, mock, iterate) as "becoming obsolete" there, and Choi's account of the pod model, where designers manage fleets of Claudes and ship in working code, points the same way. If your working style depends on that sequence, the cultural fit question is real. The team is currently ~15 designers (unverified) and being doubled, so the org structure you join may not be the one you're in six months later.

ServiceNow — Sr. Staff Product Designer, AI/ML

Trigger: Sr. Staff Product Designer (AI/ML) posting, Oct 8. High confidence.

Window: 3 days old, peak freshness. The September developments log already noted three other senior design postings at ServiceNow (Security and Risk Workflow, Customer and Partner Experience Director, SalesCRM Principal). This AI/ML role makes four concurrent senior design hires, which points to broad investment in the design org.

Snapshot: ServiceNow is a $200B+ market-cap enterprise platform that unified its AI portfolio under Otto at Knowledge 2026, merging Now Assist, the $2.85B Moveworks acquisition, and AI Experience into a single conversational front door. The broader AI platform includes AI Agents (autonomous workflow execution), AI Control Tower (governance and monitoring), and Context Engine (shared business context). As of April 2026, AI ships in every product tier by default.

What this trigger reveals: The posting describes "strategic ownership of an AI product" that "shapes large data sets to accelerate and optimize an organization's AI system" and is "early and evolving." That doesn't read like Otto (conversational) or Now Assist (feature-layer). It most closely resembles AI Control Tower or a related governance and optimization product, a tool for enterprise customers to manage and tune their AI deployments. Moderate confidence on this inference; the posting doesn't name the product.

Challenge taxonomy: Enterprise AI governance at the orchestration layer. The designer defines how enterprise customers monitor, optimize, and govern fleets of AI agents deployed across their organizations. Your Alibaba enterprise platform experience is relevant proof on the enterprise scale and data-visualization dimensions. The orchestration problem connects to your TinyFish work, though the enterprise context is different.

Landscape context: The Otto unification and the Moveworks acquisition are the largest enterprise AI platform consolidation in this cycle, and Horizontal AI hasn't covered the Otto architecture yet. What would sharpen this evaluation is knowing whether bundling AI into every tier creates a different design mandate than a standalone AI product does.

Rubric:

Verdict: Watch. Company 8, Role 13. The comp is the strongest in this cycle ($480K reported TC), and an early, evolving product inside ServiceNow's AI platform offers real scope. But the company score falls short of Act. This is a large public enterprise where design influence operates within established organizational structures, and the AI exposure is at the governance layer rather than the core product.

What would change the call: Confirmation that the product is AI Control Tower or a similarly high-visibility AI governance surface. Or a conversation showing that the "early and evolving" product has 0-to-1 characteristics and isn't an extension of existing ServiceNow analytics.