What's Actually Getting Funded in AI Agents in 2026

Eight real rounds, four themes, one pattern

AI agent funding in 2026 is not one big wave. It's four separate lanes moving at different speeds, and reading them side by side tells you more than reading any single round in isolation. Here are 8 rounds I verified directly against primary sources, not aggregator summaries, sorted by size.

Coding agents ($1.9B+ across just 2 rounds) dwarf every other lane in dollar size. Eval, security, and observability ($125M across 3 rounds) has the highest round count. Compute infra has 1 giant check. Memory has almost none, which is itself the finding, and the part worth slowing down on. If you want to see how these categories map onto specific product gaps instead of company names, browse live demand signals pulled from the same builder communities these startups came out of.

Coding agents are pulling the biggest single checks, and the field is nearly closed

If you're picking a lane purely by check size, it's coding agents, but the field already has entrenched, well-funded players. Cognition raised more than $1 billion in May 2026 at a $26 billion post-money valuation, up from a $10.2 billion valuation just 8 months earlier. Devin now runs at a $492 million annualized revenue rate and writes 89% of Cognition's own code. Its customer list spans Citi, Goldman Sachs, Mercedes-Benz, Dell, Santander, and the U.S. Army.

Anysphere, the company behind Cursor, was in talks for a $2 billion round at a $50 billion pre-money valuation by April 2026, reported by TechCrunch citing sources close to the deal. That's up from a $29.3 billion post-money valuation just 5 months earlier, in November 2025. Cursor's reported ARR at the time of those talks was around $2 billion.

Then there's 8090 Labs, which closed a smaller but still substantial $135 million Series A on June 29, 2026, led by Salesforce Ventures with Chamath Palihapitiya taking the CEO role. Its product, Software Factory, targets enterprise coding teams that need audit trails and controls a raw coding agent doesn't give them. That's the pattern inside coding agents specifically: the general-purpose coding agent slot is essentially locked up by 2 companies with $27 billion and $50 billion price tags between them. The open seats are narrower, enterprise-specific wrappers like 8090 Labs, not another general Devin competitor.

Eval, security, and observability is where the round count is highest

3 separate rounds landed in this lane inside a single 6-week window in 2026, more than any other theme in this article, which tells you investors see recurring, not one-off, demand here.

Patronus AI, founded by former Meta AI researchers, raised a $50 million Series B on June 25, 2026, led by Greenfield Partners with Datadog and Samsung both participating. Its pitch is simulated "digital worlds" that stress-test an agent's behavior before it touches a real production system.

Arcade.dev raised $60 million in a Series A on June 15, 2026, led by SYN Ventures with Morgan Stanley and Wipro as strategic investors, bringing its total funding to $72 million. Its job is proving which agent took which action on whose behalf, the kind of audit trail question that shows up the moment a company puts an agent anywhere near money or customer data.

InsightFinder raised $15 million in a Series B on April 16, 2026, led by Yu Galaxy, applying machine learning it has used for IT infrastructure monitoring since 2016 to a new problem: figuring out where and why an AI agent went wrong. 3 rounds, 3 different angles on the same underlying question, all funded within 10 weeks of each other. That density is the signal, not any single check size.

Compute and memory are the quieter infra bets, and memory is the gap

Together AI closed an $800 million Series C on July 1, 2026, led by Aramco Ventures with Nvidia, Vista Equity Partners, and Salesforce Ventures participating, landing at an $8.3 billion valuation, more than double its $3.3 billion valuation just 16 months earlier. Together AI reports over $1.15 billion in annual bookings and expects to grow its compute footprint roughly 50-fold over the next 5 years. That's the infra bet: agents need somewhere to run and someone to run inference cheaper than the frontier labs.

Memory is the opposite story. Jedify raised $24 million on June 10, 2026, to give agents structured context about a company's own business data, a narrower and more enterprise-flavored version of memory. Outside that, dedicated agent memory funding is thin. Pulling Clawsmith's own idea data on this theme right now:

get_ideas(category="ai_agent_mcp", sort="score", limit=8)

→ cross-agent-memory-mcp-server
  aggregate_score: 65,462   lead_count: 0
  "mem0 hit 58.5k GitHub stars and raised $24M Series A solving this for
  single-agent flows, but two agents working the same project still
  maintain isolated silos with no shared knowledge."

→ cross-runtime-mcp-memory-server
  aggregate_score: 22,858   lead_count: 158
  "Memory that follows you across Claude Code, Cursor, and any
  MCP-compatible runtime without per-tool integration."

Read those 2 numbers side by side. The narrower idea (memory shared BETWEEN agents, not just one agent remembering across sessions) scores 3x higher and has zero people already chasing it. The broader, more obvious idea (memory that just follows you across tools) scores lower and already has 158 leads attached, meaning a lot of builders have already spotted it and started reaching out to the people who asked for it. Same theme, opposite crowding.

How to track this yourself, with the exact queries

You don't need a paid database to reproduce any of the numbers above. 3 free searches cover almost everything a paid tracker would surface.

Site-scoped search directly on TechCrunch, updated weekly for whatever niche you're watching:

site:techcrunch.com "raises" "Series" "AI agent" 2026
site:techcrunch.com "raises" "Series" "coding agent" 2026
site:techcrunch.com "raises" "Series" "AI memory" 2026

Crunchbase News runs a dedicated AI funding tag that updates faster than most newsletters:

https://news.crunchbase.com/sections/ai/

And a standing Google Alert catches anything that slips past both of those:

"raises $" "Series" "AI agent" -site:aifunding.me -site:aifundingtracker.com

That last exclusion matters. Searches for AI funding surface a lot of aggregator and SEO-farm sites that repackage real rounds with invented framing, sometimes inventing details that were never in the original announcement. Every number in the table above traces back to TechCrunch, Bloomberg, or the company's own press release, not a roundup site. Verify the primary source before you cite a figure anywhere, especially a valuation number, which gets rounded and re-rounded as it spreads.

The threshold that separates crowded from hot but open

2 concrete thresholds, both built from what's actually in the data above rather than a vibe.

Crowded, avoid: 3 or more funded companies solving the same narrow problem inside a 6-month window. Agent security and identity crossed this line in June 2026 alone. Arcade.dev's $60 million round landed on top of named incumbents like RuntimeAI, Capsule Security, and TrustLogix, all already selling to the same enterprise fleet-operator buyer. A related but narrower idea, a lightweight kill-switch for individual developers rather than enterprise fleets, still scores 20,000 on Clawsmith with zero leads, because the funded players are all built and priced for a different buyer entirely.

Hot but open: an idea whose Clawsmith aggregate score sits above 50,000 with fewer than 10 leads attached. High score means real, sourced demand behind it. Low lead count means the people who generated that demand haven't been contacted by anyone building it yet. Cross-agent shared memory, above, is the clearest example live on the platform right now: score of 65,462, 0 leads.

The rule as a plain check, in the order I actually run it:

if funded_competitors_in_niche(window="6mo") >= 3:
    verdict = "crowded, avoid the direct pitch, look one slice narrower"
elif idea.aggregate_score > 50000 and idea.lead_count < 10:
    verdict = "hot but open, real demand, nobody's contacted these people yet"
else:
    verdict = "unclear, read the linked signals before committing"

If you want to run this check yourself before you commit to a build, pull the idea, read its linked signals for the real complaint threads behind the score, and check get_leads on it before you touch the pricing page. A high score with a fat lead list is validated but already being worked. A high score with an empty lead list is the gap this article is actually about. More funding-pattern breakdowns like this one live in the whats-getting-funded category.