AI Agent Startup Funding: October 2026 (Q4 Tracker)

Data cut-off: 4 October 2026. Q4 is open and checked weekly, and the $260.0 million above is a floor that will move. Every total on this page carries the date its research window closed, so a reader arriving in December can tell a current figure from a stale one without emailing me.

This is the running record of disclosed AI agent startup funding rounds in Q4 2026. Two rounds is two rounds, and the opening of a quarter stays an opening here. Precision about what is known, and about where the record stops, is the reason anyone cites this page. The substance of this first edition sits further down: the Q3 close moved, and it moved a lot.

The Q3 2026 funding tracker stays the per-round record for July, August and September, with a named source on every line. This page carries Q4 forward and corrects the Q3 total.

Q4 2026 rounds so far

Two disclosed rounds, $260.0 million, both dated 1 October 2026. The table below is open and gets checked weekly for the rest of the quarter.

Q4 2026 to date: 2 rounds, $260.0 million. Data through 4 October 2026. Open quarter, checked weekly.

Neither investor set is recorded here. A search summary reported Armadin's round as co-led by two well-known firms, and no primary source confirmed it, so the lead column stays out of the table. That is the rule on this page and it has cost me specificity before.

Armadin is worth one more line because of its size. A $255.5 million Series B at the start of a quarter is not a seed-stage bet on a category forming. It is a growth cheque into agent-adjacent security, and it arrived one day after Reco's $55 million growth round in the same space.

I am not publishing a median or an average cheque size for Q4 until the quarter has five disclosed rounds; the headline numbers add both automatically at that point. With two rounds that sit at $255.5 million and $4.5 million, a mean of $130 million would describe no round that actually happened. The Q3 record had 59 rounds to average across. This one has two.

November 2026 rounds

Empty by design. November has not happened as of the 4 October 2026 data cut-off on this page, so there is nothing to record. This section gets its table at the November update, built the same way the October table was: one row per disclosed round, a named source per row, and a dated total at the bottom.

If you are citing this page before then, cite the October table and the Q3 close. An empty month here records only that I have not checked November yet. Read nothing into it.

December 2026 rounds

Also empty by design, for the same reason. The December table arrives at the December update and closes the quarter. When it lands, this page will carry a Q4 2026 final close with its own shown arithmetic, in the same format as the Q3 close below.

Acquisitions and shutdowns this quarter

Money leaves the category as well as entering it, and nobody keeps a running list of the exits for agent companies specifically. This section does: every acquisition, shutdown or pivot of a company whose product is an AI agent, an agent platform, agent infrastructure or an agent-building tool, each with the announcement it came from.

As of 4 October 2026, Q4 2026 has 1 acquisition of AI agent companies, 0 shutdowns and 0 pivots on record, and no deal value disclosed by any party. The table also carries 2 deals from the last 21 days of the previous quarter, marked with the quarter, because they landed after that edition closed.

Amounts count only when the buyer or the company disclosed them; reported prices stay in the detail column and never enter a total. Data through 4 October 2026.

Two deals from the same window are left out on purpose. HCLSoftware agreed to buy Robotiq.ai, a robotic process automation platform, and LiveKit bought Loophole Labs, a virtualization company; both buyers bought them for agent work, but neither target sells an agent product. ServiceNow's purchase of Sweep and Palo Alto Networks' purchase of Console were both announced in the first days of September, outside this window, and belong to the Q3 story.

Three rounds that close out Q3

The Q3 tracker's 28 September pass stopped its September section at 24 September. Three rounds landed after that and belong to Q3. This edition records them, and the Q3 page was closed and restamped in the same batch on 4 October 2026, so both pages agree.

Three rounds, $1,075.0 million, all dated 28 to 30 September 2026 and all counted in Q3.

On Instinct's investors: TechCrunch's piece is headlined as a $1 billion Series C at a $10 billion valuation and reports the company confirming the round in a press release. A Bloomberg article from the same day reached me only as a search summary, so no investor names go in the table. The amount, the stage and the valuation are what I can stand behind.

Reco's valuation is a good example of why this page keeps valuations in a separate column. CEO Klein said the valuation has "more than doubled" since February and placed it vaguely in the "high hundreds of millions" without giving a figure. That is useful context and it is not a number. It stays in prose and it never enters a total.

Restate is the smallest of the three and the one I find most interesting, because it is the second durable-execution round in the Q3 window. Temporal Technologies raised $550 million on 14 September for the same layer of the stack. Two rounds for the plumbing that keeps a long-running agent from losing its place is a signal about where the hard engineering problems actually are.

Q3 2026 final close

Q3 2026 closes in this compilation at 59 disclosed rounds and about $6.32 billion. These totals are my own compilation of publicly disclosed rounds. They are not a market report, they are not reconciled against Crunchbase or PitchBook, and they do not claim completeness. The Q3 tracker carries the per-round source for every line that makes up the figure.

The arithmetic, shown

Here is the whole calculation, because a total nobody can reproduce is a total nobody should cite.

September: $3,209.2M at the 28 September pass, plus Instinct $1,000M, Reco $55M and Restate $20M, giving $4,284.2M. Q3 moves from $5,246.0M to $6,321.0M, about $6.32 billion.

One note on rounding, since this page gets quoted. Rounding the 28 September subtotal to $5.25 billion first and then adding $1.075 billion yields $6.325 billion. The month-level sum is $6,321.0 million. I treat the month-level sum as authoritative and quote it as about $6.32 billion.

Cognition excluded: 58 rounds and about $4.32 billion. Cognition closed on 8 September at what the company calls over $2 billion, led by Andreessen Horowitz and Accel, at a $48 billion valuation. No exact figure was published, so it is recorded at a $2 billion floor. That single round is roughly a third of the quarter, which makes the unadjusted $6.32 billion a poor input for any quarter-over-quarter comparison. Use $4.32 billion across 58 rounds when you compare Q3 to another quarter, and say which basis you used.

The four largest Q3 rounds

Each of these is already sourced on the Q3 page, line by line. Three of the four sell to developers or enterprises: coding agents in two cases and the execution layer underneath them in the third. The fourth is the consumer exception.

The four rounds above are $3,750.0 million of the $6,321.0 million Q3 total, roughly three-fifths of the money across four of the 59 rounds.

AI agent funding history, as far as this page has compiled it

This franchise has covered the period since Q2 2026, and compiled it round by round since Q3 2026. The table below covers both, and that split is the honest answer to the question people actually arrive with.

There is no 2022, 2023, 2024 or 2025 total on this page. Those years are absent because this page publishes only figures it has compiled itself against named sources, and I have not compiled them. Several third-party totals for those years exist and I could paste one in, which would make this section look complete and would make every figure on it unverifiable by the standard applied to the rest of the page. A number you cannot rebuild from a table on this site does not get printed on this site.

Every row is a Gravity compilation of publicly disclosed rounds. Each figure links back to the table it was built from. Jump to any row with its fragment, for example #stat-q3-2026.

Why Q2 2026 has no round count

The Q2 2026 edition of this tracker was written as a four-fact recap of the quarter. It names Cognition's late-May raise of over $1 billion at a $26 billion valuation, it records the valuation jump from $10.2 billion in roughly eight months against a $492 million revenue run-rate, and it describes where the mega-rounds went. It stops short of listing every disclosed round, so there is no count to carry forward.

I am leaving that gap visible. Back-filling Q2 now would mean compiling a quarter five months after it closed, from coverage that has already been reorganised by whatever happened since, and the result would be weaker than what the Q3 and Q4 tables can claim. The Q3 tracker holds the Q2 recap in full if you need the qualitative picture.

What changes across the compiled quarters

Two quarters of comparable data is thin, and I will say what it supports and stop there. Q3 2026 produced a $6.32 billion compiled total across 59 rounds, with the top four rounds accounting for roughly 59 percent of the money. That is concentration at the top of the table.

Q4 2026's first four days continue that shape at a smaller scale, with one $255.5 million round and one $4.5 million round. Whether the shape holds is the open question, and it is the thing the next few monthly updates will actually answer. For the demand side of the same period, the AI agent adoption statistics page tracks deployment with the same sourcing rules, and the AI agent pricing statistics page covers what buyers in that market actually pay.

What the pattern says

Three observations hold up against the rounds listed above. I have deliberately stopped at three, because a fourth would need a number I do not have.

Agent security is becoming its own funding category

Reco raised $55 million on 29 September and Armadin raised $255.5 million on 1 October. Back to back, straddling the quarter boundary, $310.5 million between them. TechCrunch's own framing on the Reco piece was that AI agent security startups are crowding the market, which is a notable thing for a technology publication to say about a category rather than a company.

Look at what those two companies do. Reco finds and secures the agents already running inside a company. Armadin chains security weaknesses into attack paths so a defender can test them. Both sell the same assumption: agents are already deployed, already acting, and already a surface worth attacking. That assumption is now worth nine-figure cheques.

The runtime layer keeps raising

Temporal took $550 million on 14 September and Restate took an additional $20 million on 30 September. Both sell durable execution: the infrastructure that lets a process run for hours or days, survive a crash, and resume where it stopped rather than at the beginning.

Durable execution is the right problem to be funding, unglamorous as it sounds. An agent that researches a market for six hours and then loses its place is worse than useless, because you cannot tell which half of the work is real. I have watched this failure mode from the build side and it is why the money is going to the plumbing. If you want the conceptual version of what that layer does, the AI agent architecture patterns breakdown walks through where runtime sits relative to the agent itself.

Consumer agents got a billion-dollar data point

Instinct launched an invite-only service in August 2026 and raised $1.0 billion at a $10 billion valuation on 28 September. Weeks, not years. TechCrunch framed the company as part of a new class of consumer agents that get things done instead of answering questions.

One round does not make a category, and that is the honest reading. What it does is set a reference price. Until 28 September there was no billion-dollar consumer-agent data point, and the absence of one was a reasonable argument that the money only believed in enterprise and infrastructure. That argument now needs a different piece of evidence. If you are thinking about what a consumer agent replaces in practice, the AI employee framing is the closest adjacent read.

What to watch in Q4

No numbers attached to any of these and no forecast. The current pace sits in the headline numbers; anything after the data cut-off stated there is unchecked.

• Whether agent-security funding keeps the cadence it set across the Q3 to Q4 boundary, or whether Reco and Armadin turn out to be two rounds that happened to land three days apart.
• A second consumer agent raising at a comparable scale after Instinct, which is the single cleanest test of whether that round was a category signal.
• Whether the $30 million median cheque from the Q3 compilation holds as Q4 fills in.
• Repeat raises from the durable-execution companies, and the stage they come at.
• Whether a round that gets reported as closed in Q4 turns out to be talks, which happened to Cognition for four weeks on the Q3 page.

Methodology

I compile this tracker and I check every round against a named source. I did not speak to any founder or investor for this edition, and nothing here comes from a private conversation. Where a figure is weaker than it looks, the page says so in the same sentence as the figure.

What counts as a round

A round enters a table when three things are true. The company is named, the amount is disclosed, and either a named publication reported it or the company published it itself. A company announcement, a lead investor's own post, or coverage by a major outlet all clear that bar.

An aggregator on its own falls short of that bar. A funding newsletter or a directory's daily brief is how I find a round, and the citation goes to the publication that brief credits. Where I reached a round only through an aggregator, both the brief and the publication it credits are named in Sources so you can see the full chain and judge it yourself. Three of the rounds on this page arrived that way and the chain is printed for all three.

Several things fail that test and get excluded on purpose. Reported talks are not a closed round, and the Q3 page held Cognition's reported raise out of its totals for four weeks on exactly that basis. Debt facilities are not equity rounds. A company whose product is adjacent to agents but is not an agent product gets left out, which is why a $150 million Postgres platform round announced in the same window as the Q4 rounds above does not appear here. Rumours with no primary source do not appear at all, however widely they circulate.

Valuations, ranges and floors

A valuation lives in its own column and is never added to a total. The most common error in funding coverage is a valuation repeated as a raise, usually because the headline number is bigger and reads better. When I find that error in a source, the page notes it rather than quietly dropping the round.

Where a company gives a range or an approximation, I record the floor and label it. Cognition is the clearest case on either tracker: the company says "over $2 billion" and no exact figure exists, so it sits at $2,000 million and that is the one number in the Q3 total that is deliberately too low. A floor you can see is auditable, so a floor is what the page prints.

Where a lead investor is reported but not confirmed, the table says reported. Where I could not verify an investor at all, the column stays empty. Restate's lead is in the first category and both Q4 rounds are in the second.

How the headline numbers are computed

The headline numbers are recomputed every day by a script that reads the round tables on this page and on the Q3 tracker. Nothing in that block is typed by hand. The as-of date is the latest data cut-off stated on the page, so a day with no new research does not pretend to be fresher than it is.

• Totals add disclosed amounts only. A round with an undisclosed amount counts as a round and adds nothing to the money.
• Pace is rounds divided by days elapsed in the quarter up to the cut-off. The same-day comparison counts Q3 rounds dated on or before the matching day of Q3, so day 4 of Q4 is set against 1 to 4 July.
• Median and average appear once the quarter has five disclosed rounds, because below that they describe no round that happened.
• Category share uses the Category column of the Q4 table (Security, Agent infrastructure, Coding agents, Consumer agents, Vertical agents, Agent platforms) and appears with the median.
• Exits count an acquisition when the buyer or the company announced it, including signed agreements that have not closed. A reported price stays in the detail column and never enters a total.

The full Q4 table is in data.csv, one row per round, with amounts in US dollar millions and a flag for amounts recorded as a floor.

How a correction gets logged

Corrections go in the update log with a date, and the figure that changed stays visible in that line. A reference page earns the name by keeping its own history readable.

The Q3 tracker is the proof that this matters. It was revised upward five consecutive times, always in the same direction, and 15 September 2026 alone was corrected three separate times. The third correction found Factory's $200 million round, which two earlier passes had missed on a day they had both already reviewed. The Q3 close in this edition is the sixth upward revision of that quarter. Treat any open-quarter total here as a floor, including the Q4 figure at the top.

If a round is missing, send it through contact with a dated source and it goes in on the next pass, with a line in the log saying where it came from.

Why Gravity has not raised

Gravity has raised nothing. It is bootstrapped and built in Bangalore, and I run it as the fourth company I have started after shutting three down.

That history is the reason. Each previous time, outside expectations pulled the product toward what the next deck needed and away from what the people using it needed. Bootstrapping leaves one constituency to answer to. It also means I read a funding announcement as a list of new obligations a competitor has taken on, which is probably why this page spends more words on what a round proves than on how big it was.

The product is in private alpha. If you want to see what an agent run looks like before applying, walk through a run first. For how the costs compare across the category, the cheapest AI agent platforms breakdown is dated and maintained the same way this tracker is.

FAQ

Update log

• 4 Oct 2026 Published as a live tracker. Added the headline numbers block (recomputed daily from the tables), a Category column on the Q4 table, a downloadable data.csv, and the acquisitions and shutdowns section, seeded with Salesforce and Listen Labs (29 September), ZoomInfo and DoubleO.ai (30 September) and Inworld and Ultravox (1 October).
• 4 Oct 2026 First edition. Q4 opened with 2 rounds and $260.0 million, data through 4 October 2026. Added Instinct, Reco and Restate to the Q3 close, which moves Q3 to 59 rounds and about $6.32 billion. Armadin's reported co-leads and a reported $80 million round from a separate company were both excluded for lack of a primary source, and a $150 million Postgres platform round was excluded as out of scope.

Every future pass appends a line here with its own date and its own window. If a number on this page changes, the reason it changed will be in this list.

Two sections are deliberately unbuilt at this edition. November 2026 and December 2026 are labelled empty and get their tables at their monthly updates. The funding history section starts at Q2 2026 because that is where this franchise's coverage starts, with the round-by-round compilation beginning a quarter later in Q3 2026.

Previous editions

For Indian agent startups specifically, round by round with a stated method, see the India AI agent funding tracker.

• AI agent funding tracker, Q3 2026: closed at 59 disclosed rounds and about $6.32 billion, with a source on every round. The Q2 2026 edition was folded into it and its recap lives there.

Sources

Excluded from the Q4 table and named here for the record: a reported $80 million Series A that no primary source confirmed, and a $150 million round for a Postgres platform company, which is infrastructure but not an agent product.