What the Labs Know and What a Humanoid Company Ships

The Bay Area Robotics Symposium (BARS) at UC Berkeley was held on October 30, 2026. Faculty and students from Berkeley, Stanford, UC Santa Cruz and UC Davis spent the day describing what robot learning can and cannot do. Much of the research addressed challenges posed by and applicable to humanoids.

Credit: Agility Robotics

In early October, Agility Robotics, which began trading on Nasdaq (AGLT) on November 5, 2026, presented its plans to investors. Agility leads the companies that deploy humanoid robots commercially in industrial settings, which makes it the best available benchmark for how far humanoid technology has come in practice. 

Comparing the research presented at BARS and industrial reality at Agility shows how far lab findings have traveled into products and where they have yet to arrive.

The comparison reveals a deliberate strategy: Agility builds on the diagnosis the labs have reached, sidesteps the problems they have not solved, leaves the research frontier for later, and invests more in safety than any lab. Beneath these choices lies a bet: foundation models will keep improving without Agility's help, and lasting value lies in everything around them. One line of research, agentic robotics, sharpens that bet and questions one of its pillars.

Disclosure: I hold no position in Agility Robotics and do not plan to take one.

Agility builds on the labs' diagnosis

The strongest work at BARS 2026 traced why vision-language-action models, the currently leading approach to robot intelligence, still fail. The models learn by imitating human demonstrations and copying a task without ever practicing it, and hence they never reach mastery. The models only see a narrow slice of the world because robot data remains scarce, and they falter when something new appears. And they learn mainly from images and words, which carry little information about force, so they cannot judge how hard to push, grip or insert.

Agility has built its strategy around the same diagnosis. It divides its software into two parts. "Semantic AI," built on language and vision-language-action models, decides what the robot should do. "Physical AI" decides what the robot can reliably do. The company credits foundation models with breadth and looks elsewhere for reliability. Several speakers at BARS, including Waymo and Figure AI, described the same structure: a fast reflexive layer beneath a slower reasoning layer.

The researchers and the company also agree that robots need practice. A year ago researchers proposed adding reinforcement learning to these models, and this year they showed models trained that way. Agility says its robots refine their skills through repetition, both in simulation and on customer sites.

On force, the company speaks more bluntly than most researchers. It says physical intelligence requires compliant, force-controlled hardware and that software cannot shortcut physics. Researchers are reaching the same conclusion through touch. At BARS, Berkeley's Jitendra Malik declared, "I have done 40 years of vision. The future is in tactile."

Agility sidesteps the open problems

BARS highlighted two unresolved problems: robots lack training data, and they forget. Agility answers the first with deployment, since every robot at a customer site collects data that improves the fleet. It avoids the second through its choice of work. Digit moves totes and handles material in structured, repetitive workflows. The tasks are short, so the robot needs little memory. The tasks repeat, so the robot rarely meets anything new. Choosing such work is sound engineering, but it leaves the research problem unsolved.

The same choice answers a question that Andra Keay of Silicon Valley Robotics raised in a spirited keynote: should robots be generalists or specialists? Agility answers that a robot can be both, in sequence. It describes Digit as a multipurpose worker, yet it deploys Digit today as a specialist in a narrow set of logistics and manufacturing tasks. Generality is on the roadmap, and the roadmap depends on research the labs have not finished.

Agility leaves the frontier for later

Two frontier themes from BARS do not appear in the presentation. The first is world models. Waymo showed two at work: one inside the car that predicts how a scene will unfold, and one built on Google DeepMind’s Genie 3 that generates rare and dangerous scenarios for the driving system to train on. The second is fine manipulation: several labs pushed touch sensing in that direction. Agility’s Digit 4, the robot deployed today, grasps and carries, and the company’s own capability map places assembly and fine manipulation in the future.

Agility leads the labs on safety

Safety gets more of Agility's attention than any learning method. Digit 4 works inside a fenced cell. Digit 5, planned for 2027, is designed to work near people. Agility points to a newly approved ISO safety standard and a partnership with NVIDIA on safety certification. At BARS, safety filters drew several talks but were only one theme among many. A company that must put robots beside workers ranks its problems differently from a lab.


Agentic robotics questions the data flywheel

Ken Goldberg of Berkeley gave the talk with the sharpest implications for companies like Agility. He framed robotics around two gaps: 

• The first is a data gap. Robot models train on about a year’s worth of robot data, while language models train on the equivalent of a hundred thousand years of text. Robots will not close that gap by collecting more.
• The second is a revenue gap. Traditional programming is reliable but narrow, and learned models are general but unreliable, so neither delivers dependable, flexible work at scale.

Goldberg’s lab proposes a third path: let frontier coding agents write the robot program. Several agents each refine one part of it, such as perception, motion planning or grasping, and rehearse the program in simulation before it reaches the robot. An AI model can even build that simulation from a phone video of the real scene. On a grocery-picking task in simulation, which is varied but structured work much like Digit’s, the approach succeeded 97% of the time; a leading vision-language-action model succeeded 20% of the time. Goldberg credited four of his students for the work: Eric Kaiyuan Chen, Shuangyu Xie, Uksang Yoo and Xinyue Zhu.

This approach bridges the two cultures of robotics, engineered control and learned models, much as Agility’s own split between physical and semantic AI does. But it reaches the bridge by a different route. If agents can write reliable robot programs without large sets of demonstrations, then a fleet’s data becomes a weaker advantage than its skill library, its simulations and its hardware. Agility’s presentation treats deployment data as a central asset and does not address that possibility.

The bet behind the strategy

Agility's choices rest on a view of how quickly AI models age. On a recent 20VC podcast, Rory O'Driscoll of Scale Venture Partners cited Anthropic's S-1 filing, which shows that Anthropic earns about 80% of its revenue from models less than six months old. He estimated that training consumes most of the company's compute. Each model therefore has only a short window to repay its cost.

If robot models age as quickly, a robot company cannot rest on its model. Agility treats foundation models as inputs that will keep improving, and it invests in things that last longer: hardware that handles force, a safety case that regulators accept, a factory, deployments that generate data, and software that runs fleets at customer sites. Agentic robotics extends the same logic: When frontier coding models improve, the robot programs they write improve with them at no cost to the robot company. It also suggests that, of these assets, data may hold its value the least.

The work presented at BARS supports that ordering of priorities. It also shows how much of manipulation remains unsolved.

Robot Foundation Models Work. What Breaks Next?

Notes from the Bay Area Robotics Symposium, 2025 vs. 2026

On October 30 I attended the Bay Area Robotics Symposium (BARS) for the second year. Faculty from UC Berkeley, Stanford, UC Santa Cruz and UC Davis gave dozens of short talks, and industry and student speakers added spotlights.

My notes from the two years show a shift. In 2025 speakers asked how to build robot foundation models. This year they asked why the models still fail.

What stayed the same

Robots still lack data. Language models train on the text of the internet in billions of hours; a leading robotic model two years ago was trained on ten thousand hours of robotic data. To accelerate, labs edit human videos to resemble robot demonstrations, build cheap teleoperation rigs, rebuild real scenes in simulation and generate synthetic data.

Robots still forget. Models lose track of earlier steps in a long task, and they lose old skills when they learn new ones. Researchers proposed retrieving only the key moments of a task and anchoring old skills in memory. No one claimed a solution.

Simulators remain hard to use. Researchers praised the leading physics engines and again complained about them. Contact, soft materials and small forces still resist accurate simulation.

What was new

Researchers now agree on why vision-language-action models fail. A model reaches about 80% success on the tasks it trained on; the last 20% remain the hard part. Speakers named three causes: The models do not master tasks, they do not generalize beyond their training, and they do not apply force precisely. Most of this year's presentations addressed one of the three.

Mastery. Last year researchers proposed adding reinforcement learning to these models. This year they showed models trained that way, and those models perform tasks more reliably. One lab simplified the training by stating the reward in plain language instead of building a separate reward model.

Generalization. Researchers showed models that reason in images and words before they move. An autonomous-driving company described two of them: One world model runs in the car and predicts how the scene will unfold. The other, built on a video model, generates unusual scenarios on which the driving system trains. Some researchers even call this latter language model an implicit world model, but they do not predict failure states of a scene from actions.

Force. Touch moved to the center. Last year tactile work meant mostly new robot hands. This year several researchers argued that the next advances will come from touch, not vision. Speakers also named hardware limits: actuators overheat, and designers must choose between three fingers for precision and four for force.

The problem as a whole. Two developments address all three causes at once. First, companies in driving and humanoids now split control into a fast reflexive layer and a slower reasoning layer, so that one layer supplies reliability and the other breadth. One humanoid maker even uses three layers. Second, coding agents entered robotics. Researchers asked whether AI agents can write robot software. Long programs confuse the agents, so some groups break the program into a graph of small parts that agents build and revise one at a time. Because the agents need no demonstrations, the approach sidesteps the data shortage.

Andra Keay of Silicon Valley Robotics gave a spirited afternoon keynote. She asked what is really happening with humanoids, whether robots should be generalists or specialists, and what animals can teach us about the choice. 

This year's research presentations supported answers to each choice. Foundation models push toward generalists, while layered designs and coding agents favor reliable specialists.

Back What Every Winner Needs: A Defense Playbook for Angel Investors

The war in Ukraine has made unmanned vehicles central to modern combat, and the cycle of measure and countermeasure now runs in months and weeks. No outsider can follow that pace, and no early-stage investor can predict which vehicle will win. An investor can, however, back what every winner will need. That idea leads to three questions an early stage investor should ask of defense startups: Where does it play? Is the demand real? Who will pay for it to scale?

The Market Rewards Discipline

Government demand is real but narrow. The Pentagon's Drone Dominance program, an 18-month, $1.1B effort, aims to buy hundreds of thousands of one-way attack drones by the end of 2027. Yet its first round split a $150M order among 11 companies, and its second drew 48 companies with about 78 designs.

Much of this has been financed by expensive equity capital. Defense and adjacent startups raised $35.6B in the first half of 2026, and early-stage companies reportedly raised at 17 to 50 times revenue. Anduril's CEO Brian Schimpf calls the market ‘in a bit of a bubble’ and warns that companies burning cash without production contracts will face down rounds.

Where Should an Early Stage Investor Play?

Remove a bottleneck. Back companies that remove a bottleneck rather than add one more platform. Training and simulation tools for autonomous vehicle operators are a case in point, because operator skill limits this type of warfare more than hardware does. Supply-chain software also checks that box since every manufacturer in the Pentagon's drone program must prove that motors and batteries come from outside China.

Invest in a layer, not a whole system. A stable layer, such as motors, batteries, magnets and sensors, sells no matter which manufacturer wins. A fast layer, such as jam-resistant communications, navigation without GPS, or swarm control, keeps pace with the front and sells for every vehicle. A seed-stage drone airframe, by contrast, competes as one more platform, and the next iteration of the war will make its design obsolete.

Require an advantage rivals cannot copy. A prime contractor or a better-funded rival must not be able to rebuild the product in a quarter. The advantage can be proprietary data, access to classified networks, a security authorization (ATO), or a design proven in combat, such as Ukrainian technology licensed or co-produced in the United States. Thin "defense AI" companies will fail this test.

Prefer dual use. A commercial market will carry the company when defense budgets swing. Navigation without GPS also serves agriculture and logistics; drone motors and batteries also serve industrial robots.

Is the Demand Real?

Require contracts with money and a buyer. Real traction means an awarded contract with funding behind it, a named buyer, and a path to production. A prototype contract (OTA) counts when a program office has committed to the follow-on and has named the buyer. A finalist slot, a zero-dollar CRADA or an SBIR Phase I shows interest, but it does not count as traction.

Require a funded path to a program of record. The best early-stage companies keep their burn low until a program funds them, so each equity round buys progress rather than survival. Ask the founder to name the program, the budget line and the champion who will carry the company there, and to show how long current cash lasts.

Value the company on appropriated money and price the round on contracts Congress has funded and program offices have awarded. The FY2027 budget request seeks $54.6B for the Defense Autonomous Warfare Group, up from about $225M in FY2026, but Congress has not yet appropriated it. But a budget request is only an option the company may win, not revenue it can count today.

Who Pays to Scale?

Capital-intensive hardware companies raise round after round before revenue is generated, and each round dilutes the investors who came in first. An angel's equity should therefore fund design, certification and early sales, and someone else should pay for the factory. Software companies meet this rule outright. Hardware companies meet it when one of two sources finances production.

Look for a contract that lenders will finance. US government receivables make strong collateral. Lenders finance purchase orders and receivables, and the Assignment of Claims Act lets a contractor assign government payments to a lender. Machine tools and production lines with resale value can also carry equipment loans.

Look for government capital. The Office of Strategic Capital lends to companies in critical supply chains, and Defense Production Act Title III awards fund new production capacity.

Back What Every Winner Needs

The developments in an actual conflict change faster than any outsider can follow. An angel trying to pick the autonomous vehicle that wins is akin to playing the lottery. Every autonomous vehicle, though, will need operators who can move it, components that pass the sourcing rules, software that survives jamming, and a factory that someone else but the angel paid for.

The Pentagon Has Learned From 1928. What Can It Learn From 1917?

In "Learning to Fly Again," I described how the United States entered World War One three years behind in the air, and how it caught up. On October 19, 2026, the Pentagon issued a memo and Owen West, who leads Project Agincourt, subsequently laid out how the department will build an Autonomous Warfare Command. His memo deserves a close reading against the record of 1917.

The Pentagon Remembers 1928

West opens by recalling 1928 when Secretary of War Dwight Davis ordered the Army to test a unified mechanized force of tanks, armored trucks and motorized infantry. The Army soon disbanded that unified force and split mechanization among the existing branches as add-ons. It took the German blitzkrieg of 1940 to force the creation of the US Armored Force.

Agincourt aims to avoid repeating 1928. It will stand up a four-star Autonomous Warfare Command (AUTOWARCOM) by 1 October 2027. The command's commander will have the authority to "drive the largest institutional change since the interwar years." West states the goal plainly: "The goal is not a particular drone. It is a system built to adapt at scale."

Where The Memo Echoes 1917

On four points, the memo reads as if its authors had studied the American air effort of 1917–18.

Asking for evidence. In 1917, planners promised to darken the skies over Germany, and the causes of that ultimately failed promise ended up being the subject of the Hughes investigation. West writes that new systems "will not be declared ready on paper." Competitions, exercises and combat performance will decide which systems scale. Every auditable dollar must lower the cost per kill or the risk to the force. The memo will measure success by how long a fielded solution takes to solve an operational problem.

Learning at the edge. In 1917, the Bolling Mission went to the front and asked the people who flew and maintained the aircraft which types to build. Agincourt pairs companies directly with operational units and calls those units "incubators."

Valuing the operator. American pilots were trained in French and British schools and learned from the American veterans of the Lafayette Escadrille. West highlights "an unmatched NCO corps" as one of two American strengths.

Building an institution. NACA, founded in 1915, mattered more than any single aircraft of 1918. AUTOWARCOM aims to make adaptation itself the main product of a permanent command.

1917 Asks Three More Questions

Who will build at scale? The Liberty engine succeeded because Detroit could build a stable, standardized component by the thousands, and because Allied orders had already built American capacity.
The Agincourt memo says nothing about industrial capacity: A command that cuts systems quickly after each competition protects the taxpayer, but will suppliers build factories for a demand that can disappear the next quarter?

Where is the front? In 1917, the United States learned at the front of a war it had not yet joined. The Agincourt memo relies on competitions, exercises and combat performance, yet never mentions Ukraine, where combat performance is measured every day.

Whose designs will American forces use? Most aircraft the American Expeditionary Forces received in 1918 were superior French designs and necessity had overcome national pride. Will American forces adopt Ukrainian designs where those designs outperform domestic programs?

The Clock Is Running

West quotes the Chairman of the Joint Chiefs of Staff General Caine: "If not us, who? If not now, when?" 

For investors, the questions are of  a different nature. In 1917, Detroit built the Liberty engine, and the front left the DH-4 behind. In today’s context, the questions would be “Which companies build what every winner will need in this conflict and the next? And who pays to scale them while they wait for a buyer?”

In my next post, I will turn to what these questions mean for early-stage investors in defense, intelligence and aerospace.

Learning to Fly Again: Arming for a War You Are Not Fighting

Russia invaded Ukraine on 24 February 2022, 1,707 days ago at the time of writing. This war has already lasted 140 days longer than World War One.

The United States is not a combatant in Ukraine. It also did not fight in World War One until April 1917, almost three years into the war. Until then, US banks lent the Allies billions of dollars and American factories filled their orders. A nation that watches a war from outside finds it hard to know where to invest, because the combatants change their weapons faster than an outsider can follow. The United States had to learn this in 1917, and its armed forces face the same challenge today.

America started three years behind

Since 1914, Europe's air forces had moved through several generations of fighters and had split air power into pursuit, reconnaissance, bombing and ground attack. When Congress declared war on 6 April 1917, the Army's air arm was a section of the Signal Corps. It owned fewer than 300 aircraft, none fit to fight over the Western Front, and a few dozen qualified pilots. It had no combat-tested design, no military engine in volume production, and no doctrine.

The United States did have assets Europe lacked: a vast automotive industry, money, and young men eager to fly. Allied orders had already expanded its munitions, steel and machine-tool capacity. Congress had also laid institutional groundwork: It created the National Advisory Committee for Aeronautics (NACA) in 1915, and the National Research Council followed in 1916. Both shaped American aviation for decades.

Rapid innovation outpaced America's late start In July 1917 Congress appropriated $640 million for aviation, then the largest single appropriation in American history. Planners promised more than 20,000 aircraft.

To reach that volume, the Army chose the British DH-4 two-seater,  an airplane that had first flown a year back in August 1916, fitted it with the American Liberty engine, and tooled up.
The front moved on while the factories retooled and the first American-built DH-4s reached France only in mid-1918. Crews judged the type outdated and called it the “flaming coffin” because its fuel tank caught fire easily. Only about 1,200 had crossed the Atlantic when Germany collapsed. Most of the American-built aircraft never left the United States.

America caught up by learning, borrowing and building

The United States still closed the gap within 18 months.

• It learned from the front. In June 1917 the Bolling Mission went to France, Britain and Italy and asked the people who flew and maintained aircraft which types to build. In the end, most of the roughly 6,000 aircraft the American Expeditionary Forces received were French, among them the SPAD XIII and the Nieuport 28. National pride gave way to necessity and effectiveness.

• It mass-produced what industry did best. Engineers designed the Liberty L-12 engine in about a week in May 1917. Detroit knew how to make a standardized, high-volume component, and Packard, Ford, Lincoln and others built more than 13,000 by the armistice. 

• It trained with the allies. American pilots learned at French and British schools and at the American center at Issoudun. Veterans of the Lafayette Escadrille, Americans who had flown for France since 1916, formed an experienced cadre.
  
  

Ukraine innovates in weeks

In Ukraine, measure and countermeasure follow each other within weeks. Cheap first-person-view drones made armor vulnerable. Electronic jamming answered them, and fiber-optic guided drones answered the jamming. Interceptor drones now hunt long-range attack drones. 

Both sides now build millions of drones a year, by their own accounts, mostly from commercial parts. Software, radio links, seekers and operator skill decide the outcome, and all of them change faster than an airframe.

US acquisition runs on a slower clock. A major program takes years to move from requirement to fielding. The Pentagon's Replicator initiative, launched in 2023, aimed to field thousands of attritable systems within two years, a pace the department considered fast. In Ukraine, two years span many generations of drone design.


Learnings are easy to state but difficult to act on

The lessons for today’s forces are well known and will not be repeated here. Many of them threaten the interests of the established defense industry, which explains why they are easier to state than to act on.  In my next posts, I will take a look at Project Agincourt and then turn to what this means for early stage investors in defense, intelligence and aerospace.

Same Angel Investor, Different Advice: Changes From Pre-Seed To Series A

The last piece was about how an angel’s posture changes from an informal relationship in the SAFE-and-no-board years to a more formal, renegotiated role once a board is constituted at Series A. This one is about something narrower and, in the day-to-day, more useful: what the content of good advice actually is at each transition, and why giving seed-stage advice to a pre-seed company or Series A advice to a seed company is a failure mode.

Disclaimer: Most companies raise on SAFEs for both pre-seed and seed, sometimes stacking several SAFEs across rounds that blur into each other with no priced round to mark a clean line. That ambiguity is exactly why the advice has to be anchored in what the company is actually proving at a given moment and not on the label of the round.

Pre-seed to seed: Underwriting whether the problem is real

At this stage, the founder's job is to prove the problem is real and to find signal fast. The central risk an active angel should be watching for is drift into vanity activity that looks like progress but isn't evidence of anything.

Four things are worth pressure-testing, and none of them are about the business model:

1. Founder-market fit and execution cadence
   Is the team learning quickly and shipping against a tight loop, or spinning on features nobody actually asked for? The question is whether the business is converting into signal.

2. Founder coachability, and specifically how they handle bad news
   At this stage the investor is underwriting the person even more than the plan. The plan will likely change several times before it's right. How a founder reacts the first time the investor tells them something they don't want to hear is a better predictor of the next eighteen months than anything in the deck.

3. Burn versus runway to the next real milestone
   Runway to a milestone that will justify the next raise is a forecast, and it's the number that determines whether the founder raises from strength or from desperation. 

4. Signs of premature scaling
   Hiring, spend, or a process built for a company that doesn't exist yet are the earliest version of a mistake that gets much more expensive to unwind after the next round.
   
   

Seed to Series A: Underwriting whether it's repeatable

By seed, the problem-is-real question should have mostly been answered. What Series A investors are underwriting is whether growth is repeatable and not just dependent on the founder hustle, and whether the organization is starting to look like something that can run without the founder personally closing every deal.

1. Unit economics and repeatability
   Growth that depends on the founder's personal charisma in every sales call isn't a business yet; it is a founder with a good pitch. 

2. Customer concentration and retention
   Is revenue durable, or is it resting on two or three accounts that could walk and take the growth story with them?

3. Organizational readiness
   Series A usually demands the founder shift from doer to builder-of-systems — from closing every deal personally to managing managers who close deals. 

4. Narrative discipline for fundraising
   Can the founder now tell a consistent, evidence-backed growth story to investors who weren't there for the early pivots and don't have the relationship context an angel does?
   
   

Installing the habit matters as much as naming the problem

There's a window of opportunity angel investors should use deliberately: the best moment to install the governance habits the next stage will require is right after each financing closes, and before the next fire drill starts while the founder still has bandwidth and goodwill for a process that doesn't yet feel urgent.

How that advice and governance get delivered matters as much as what gets delivered. The angel's actual edge, across both transitions, isn't reading a different checklist than the founder. It's knowing which item on that checklist is the one worth asking about this quarter. 

The socratic method of asking questions is a time proven method that lets the founder arrive at the concern themselves and not let the angel declare a problem. Founders own conclusions they reach on their own.

Your Angel Mandate Has an Expiration Date. Here's How to Renegotiate It

Zoom out from any single company, and an angel's career follows a fairly consistent shape in three time-bound phases for any invested cohort.

Phase	Typical duration	What's actually happening
Active deployment	Year 0 - 4	Capital goes out. A diversified angel typically writes 15–30 checks in this window.
Management & follow-on	Year 0 - 6 years	Focus shifts to mentoring existing founders, monitoring the portfolio, and deciding where to defend a position with follow-on capital.
Waiting & liquidity	Year 3 to 15+	Most of the calendar. Angel-backed companies take roughly 9–11 years, on average, to reach an acquisition or IPO, so even a short active window is followed by a long one spent simply waiting for an outcome.

Two structural forces compress the active window. One is capital: angels invest personal money, and once the "high-risk" allocation is deployed, new checks wait on prior ones returning capital. The other is fatigue: with 60 - 70% of individual investments returning zero and full-cycle exits often a decade out,  many solo angels burn out and step back after ten to fifteen years. 

Somewhere around Series A, the governance picture most angels have been informally operating on disappears and a real one replaces it. It's worth spelling out what actually changes, because the instinct to keep doing what worked until then is usually the wrong one.

A few things happen at once. The company converts from SAFEs to preferred stock and a board gets formally constituted. Typically the lead Series A investor takes a seat, sometimes a second institutional investor, hopefully an independent director. Board members now carry fiduciary duties such as care and loyalty to the company and its shareholders as a whole. This is a materially different obligation than an angel casually advising a founder. Information starts flowing through formal channels: board decks, monthly or quarterly board packets, sometimes committees for compensation or audit. And the founder's time, which used to be available for an unscheduled call with a helpful early check-writer, now is formally claimed by the people with a board seat and a fiduciary reason to use it.

Most angels do not end up with a seat at that table. Check sizes at seed are usually too small relative to the Series A lead to justify one, and a crowded cap table of individual board seats is exactly what a new institutional investor will resist. So the realistic question isn't "how do I get a board seat", it is how to define a role that still adds value. In practice, that tends to land in one of three places:

• A board observer seat, negotiated into the side letter at the seed round rather than requested after the fact. It carries forward, gives visibility into board-level decisions without a fiduciary vote, and is far easier to secure before the cap table gets competitive.
  
  

• A formal advisory agreement, with an explicitly scoped mandate and sometimes a modest equity refresh. This is the mechanism by which the "audit your superpower" boundary from the seed stage gets written down rather than left as an informal understanding.
  
  

• A narrower Strategic Connector role, where the angel steps back from day-to-day operating involvement because the company now has real executives and a real board doing that job, and the highest-value contribution left is network access when it matters: the enterprise intro, the next-round investor, the exec hire.

This is more of a recalibration of the role than a downgrade. The Operational Co-Pilot role that was genuinely useful when a company had no VP of Sales and no board is often actively unhelpful once it has both. The angels who navigate this well are the ones who treat the arrival of a board as a scheduled renegotiation of their own mandate and not a demotion.

These transitions happen because the angel decided, at each stage, what role was actually useful and asked for it. The angels who stay relevant longest are the ones who keep renegotiating the mandate as the company's governance catches up with its ambition.