How Much Agents Are Actually Making?

rep_hq avatar
REP
@rep_hqMar 26, 2026

The Spaces convened builders and ecosystem leads to examine whether on-chain AI agents can sustain real businesses after paying inference, gas, insurance, and acquisition costs. Danny (Rep) framed the trust and reputation needs for agent coordination; Danny from Banker detailed a transparent LLM gateway, safe wallet architecture, and agents’ revenue beyond trading fees. Umer (Virtuals) mapped demand across human-to-agent, agent-to-agent, and personal agents orchestrating many, stressing discoverability, deterministic outputs, and an AGDP incentive flywheel. Lorenzo (Venice) described tokenized access (DM: $1/day API credit) and agent SEO via Cloudflare’s markdown, which, along with OpenClaw, drove API adoption. Sid (Ask Gina) highlighted systematic trading strategies as a sustainable revenue source, with agents putting idle capital to work. Sam Sisana (Base) and Quickly emphasized identity, credibility, and verifying agent autonomy (“sign in with agent”). Examples included Felix’s autonomous marketplace and services ($86k in 30 days) and Austin Griffith’s surge in agent-built app users. Participants forecast agents becoming invisible orchestration layers (agentic requests supplanting APIs), token-aligned distribution remaining key, and a rising need for robust agent reputation as the ecosystem accelerates.

Agents, Revenue, and Unit Economics in the Agentic Economy — Twitter Spaces Summary

Participants and Roles

  • Danny (Host; Rep): Head of Growth at Rep (portable trust/reputation layer for the agentic economy)
  • Danny (Banker): Co-founder/lead at Banker (agent wallet + agent ecosystem; LLM gateway and skills)
  • Umer (Virtuals): Core team at Virtuals (ACP, AGDP, ECF; distribution and discoverability for agent commerce)
  • Lorenzo (Venice): Comms/no‑code background; Venice tokens VVV & DM; model/API provider and agent-friendly payments
  • Sid (Ask Gina): Co-founder of Ask Gina; early on-chain agent builder; experiments with model performance and on-chain integrations
  • “Quickly” (builder): Working on identity/verification (e.g., sign‑in‑with‑agent); focused on credibility and autonomy verification; healthcare professional
  • Sam Sisana (Base): Leads EMEA & India at Base; supports AI/agent builders on Base

Framing: What’s the real business model after costs?

  • Host Danny (Rep) set the core question: Beyond hype and screenshots, after paying for insurance, gas, inference, and customer acquisition, is there a sustainable dollar left for builders if we remove subsidies? He notes the rails exist (identities, compute—including private/decentralized compute—distribution, inference), agent–agent payments are up, and the agentic economy is visibly accelerating. Rep focuses on portable, cross‑platform reputation to enable trust and coordination among agents and humans.

Defining autonomy and verifying “real” agents

  • “Quickly” raised the need to distinguish fully autonomous agents from human-driven workflows (human-in-the-loop vs. hands-off). Their immediate focus is verifying that revenue and on-chain actions are indeed performed by agents (e.g., sign‑in‑with‑agent), then assessing autonomy level.
  • Banker Danny emphasized the continuum, not a binary: many agents operate with managed autonomy (e.g., check-ins for key actions), similar to team management. Key question: what proportion of decisions/actions is agent-led vs. human-approved?
  • Transparency signals:
    • Banker’s agent pages show both on-chain revenue/fees and LLM usage via their LLM Gateway—useful proxy for “how much work” the agent is actually doing.
    • Documentation of autonomy settings (e.g., required approvals) and telemetry can help prove provenance of agent actions.

Demand and distribution: where usage is coming from now

  • Umer (Virtuals) described three layers of evolution in demand on ACP (Virtuals’ Agent Commerce Protocol):
    1. Human → Agent: Early phase (e.g., Butler) focused on human trials and improving inference/deterministic outputs.
    2. Agent → Agent: With the rise of “Open Claw” (community agent framework) usage, ACP is seeing significant agent-to-agent calls. Example: HiFury agent, once a “ghost town” for humans, now sees consistent machine callers because it provides reliable outcomes, clear documentation, and safety checks (e.g., assessing whether a counterparty agent/wallet is safe).
    3. Personal Agent → Many Agents: Users run a personal assistant that orchestrates multiple agents to complete tasks.
  • Agentic discoverability is different from Web2/SaaS:
    • Clear, machine-readable documentation, deterministic outputs, safety and reliability, and “agentic SEO” are decisive. Agents must be easily parseable and trustworthy to other agents.
  • Lorenzo (Venice) on being discoverable by agents:
    • Venice built agent-friendly payment rails (VVV, DM tokens) so agents can pay natively on-chain.
    • Cloudflare’s “markdown for agents” integration makes docs/websites more machine-discoverable.
    • Early participation in “Open Claw” and adding Venice as a model/API provider led to a hockey-stick increase in API adoption once discoverability + agent tooling + distribution aligned.
  • “Quickly” on product focus: Build for agents first—optimize machine readability, speed, and APIs over human aesthetics. Human-facing UI can attract attention but agents need clean capabilities, observability, and fast integrations.

Building for agents vs. building for humans: what’s working

  • Banker Danny outlined Banker’s “agents-first” architecture:
    • Wallets for agents: Never give an agent a private key; use sharded/managed wallets or encrypted secret vaults (e.g., Coinbase approaches) so agents can act on-chain without key exposure.
    • Guardrails: A year-plus of hardening natural-language-to-on-chain transaction translation with robust protections against hallucinations and unsafe intent execution.
  • Revenue-leading agent examples (from Banker’s ecosystem):
    • Kelly: Pushed 12 iOS apps from ideation to App Store, showing end‑to‑end app factory capability.
    • Felix: Generated ~$86k revenue in 30 days by:
      • Launching Claw Mart (a marketplace built autonomously)
      • Selling Felix Craft (on-demand agent building services for humans)
    • Claude (Austin Griffith’s agent): Massive usage uptick vs. his five prior years of on-chain app development—attributed to token-aligned early users and attention economy synergy.
  • Banker Danny’s thesis:
    • Many top agents today build for humans (replacing or augmenting human labor), while others fill gaps for agents (e.g., discoverability marketplaces).
    • Tokenization is a powerful bootstrap alternative to venture funding—pays inference, aligns early users as holders/advocates, and compresses the PMF loop to days/weeks.
    • Optimizing unit costs (LLM/inference) is secondary once PMF is found; the first priority is growth and distribution, then operational tuning.

Beyond fees: concrete agent revenue models being validated

  • Services for humans:
    • App factories (e.g., Kelly), bespoke agent development (Felix Craft), verticalized “AI as a Service” offerings.
  • Marketplaces for agents and humans:
    • Discoverability markets (e.g., Claw Mart), listing skills/capabilities other agents can call; revenue via fees, subscriptions, or usage-based pricing.
  • Trading and on-chain strategies:
    • Sid (Ask Gina): AI lets anyone deploy systematic and continuous trading strategies across chains/markets with natural language, turning capital into an “always-on” liquid hedge fund. Agents journal, learn, and refine strategies (RL-style), finding sustainable edges in inefficient or uncorrelated markets (e.g., DeFi/arbitrage/prediction).
    • Banker Danny extrapolated: In the agent-native world, idle balances will be continuously optimized—agents will put unused funds to work automatically (e.g., simple yield, DeFi, or by outsourcing to specialized trading agents). This creates a background economy of capital utilization and agent-to-agent service markets.
  • Token-aligned revenue and value accrual:
    • Umer (Virtuals): Tokenization bootstraps early development (replacing VC), then—crucially—token design should tie usage to value accrual, driving a flywheel of demand. Example from Virtuals ecosystem: Rapple achieved 90k users and ~300M transaction-equivalents (activity proxy) within ~2 months via strong token design + usage.
    • Distribution is king: Virtuals focuses ACP on discoverability and demand aggregation and is moving to open-source indexing so any agent on the internet can be found/utilized (agentic SEO for the open web), plus AGDP incentives (up to $1M/month) to spur quality supply.

Managing costs: inference budgets and automatic payments

  • Lorenzo (Venice) on cost predictability:
    • DM token provides ~$1/day in API credits per token—gives builders a predictable inference budget and simplifies planning.
  • Banker’s LLM Gateway:
    • Agents pay for LLM usage automatically from revenue/fees; agents’ LLM consumption is made transparent on their public pages.
  • Stack realities (Sid):
    • On-chain costs (gas, bridging) and AI inference costs both matter. Builders are evaluating model performance for action reliability and cost profiles, optimizing across the full stack.

Trust, safety, and identity: the coordination layer

  • Host Danny (Rep) emphasized the need for a portable, verifiable, cross-platform reputation so agents can trust other agents and choose safe counterparties—especially as agent–agent payments grow and malicious/sophisticated behaviors rise.
  • “Quickly” focused on verifying provenance (are agents truly acting?) and on credibility signals.
  • Virtuals highlighted that agent safety checks (e.g., via agents like HiFury) are becoming a de facto prerequisite for machine-to-machine commerce.

Predictions and near-term outlook (through 2027)

  • Host Danny (Rep): Expect an “agentic summer” inside a niche where agents yield a clear efficiency step-change; once tooling and composability click, agent–agent activity will surge. Coordination and reputation will be decisive.
  • Lorenzo (Venice): Hopes agents ultimately reduce screen time—paradoxically everyone is more online now to build/learn; aims for a world where agents handle the busywork while humans connect IRL.
  • Sid (Ask Gina): There’s an activation hump; after your first autonomous recipe runs, it’s a “breath of fresh air.” By year-end, many more users will have meaningful automations freeing up capacity.
  • Sam Sisana (Base): Two winning directions—deep tech or “anti‑AI” human-first experiences. The crypto/AI bubble is small; mainstream adoption is early. Window of opportunity to build foundational value before mass adoption.
  • Umer (Virtuals):
    • Multi-agent orchestration becomes default: one task → agent brokers 5+ specialist agents that coordinate behind the scenes.
    • Backend shift from traditional APIs to “agentic requests” (software hiring services from agents directly). Discovery and integration simplify when agents can find and invoke each other without manual API key wrangling.
  • Banker Danny: The shift to an agentic internet (potentially “without APIs” as we know them) will change how software integrates; cites a recent XMPT founder discussion on a gentic architectures superseding classical API patterns.

Practical guidance for builders

  • Build for agents first:
    • Machine-readable docs; explicit capability schemas; deterministic outputs; clear, stable interfaces and SLAs.
    • Transparent autonomy settings and LLM usage telemetry.
  • Prioritize security primitives:
    • Managed/sharded wallets, never raw private keys in agent context; guardrails for NL→on-chain safety.
  • Design distribution in:
    • Treat “agentic SEO” as a core competency: Cloudflare agent-markdown, structured docs, examples.
    • List in agent marketplaces; integrate popular frameworks (e.g., Open Claw) and skills ecosystems.
  • Monetize beyond trading fees:
    • Services for humans (app factories, custom agents) and for agents (specialist microservices, discovery markets, safety attestations).
    • Consider token design that aligns usage with value accrual; use tokens to bootstrap inference and community while you iterate to PMF.
  • Optimize costs progressively:
    • Use predictable inference budgets (e.g., DM) and gateways that auto-charge from revenue. Once PMF is clear, swap/ensemble models for better unit economics.
  • Bake in reputation/trust:
    • Integrate portable reputation layers (e.g., Rep), safety agents, and verifiable identity (e.g., sign‑in‑with‑agent) to enable permissionless agent‑to‑agent commerce.

Representative metrics and examples cited

  • Felix: ~$86,000 revenue in 30 days via Claw Mart and Felix Craft.
  • Kelly: 12 iOS apps built and shipped autonomously.
  • Claude (Austin Griffith): More users in weeks than his prior five years of on-chain app usage.
  • Venice DM: ~$1/day API credit per token for inference budgeting.
  • Virtuals AGDP: Incentives up to ~$1M/month to spur agent supply.
  • Virtuals ecosystem (e.g., Rapple): ~90,000 users and ~300M transaction‑equivalents within ~2 months (activity proxy).

Closing

  • The group broadly agrees agents are already generating real revenue in multiple ways—and more importantly, new economic patterns are emerging: agent-to-agent commerce, capital never sitting idle, and token-aligned communities that compress the path to PMF.
  • The biggest near-term levers are distribution and trust: be discoverable to agents, and make it safe for agents to transact with you.
  • Host Danny (Rep) closed by inviting listeners to follow the guests (Venice, Banker, Gina, Virtuals, Base) and to claim a Base‑app on-chain achievement for attending; Rep hinted at upcoming releases tied to users’ on-chain agentic activity and verified participation.