Agent companies are valued in the billions, and their agents ship without a way to prove what they did. The proof layer has to be independent, and independence cannot be retrofitted. This page is the thesis, written down.
The running of agents is commoditizing in the open: reference loops are free, the reversible-execution substrate is open source, and every serious lab ships a harness. At the same time, capital prices the runner layer in the billions. In July 2026 an agent lab with no disclosed revenue and no governance story was in talks at a reported $1.5B valuation.
The adoption side moves just as fast. The same month, a major technology company publicly described embedding AI engineers next to domain experts, shipping working agents straight into finance, legal, and HR, and told the world to copy the method. The playbook ends at ship. No record, no independent review. Spegling is where the shipped agent runs.
The money already lands on the record side. The company behind a free agent framework with tens of millions of monthly downloads gives the framework away; its billion-dollar valuation stands on the commercial product, which sells traces and evaluations of agent runs. That is telemetry: useful for debugging, silent in a dispute. No sealed ledger, no reviewer from outside the vendor's family, no record the customer owns. The record layer monetizes at unicorn scale, and the governance seat above it stays open.
The vendors describe the gap themselves. In July 2026 the creator of one of the most-used coding agents published an adoption ladder for organisations: each step up is earned by automating verification a human still does by hand, and the named trap is scaling agent count before the verification loop has earned trust. The ladder's trust loop is self-verification: tests, builds, and review tooling from the same vendor's family. What it does not contain is a durable ledger or a reviewer from outside that family, which is what makes trust showable to someone who was not in the room.
All of it points the same way. Capital prices the runners, the playbook mass-produces them, and every one shipped increases demand for the one thing none of them include: an independent answer to what the agents actually did. The loop is what you demo. The ledger is what you defend.
Last verified: July 2026.
A model vendor will never route review of its own output to a competitor's model, and will never publish the failure map of its own models. Cross-family review and an independent record are structurally unavailable to every well-funded player in the stack. The incentives forbid it.
The cloud platforms host every vendor's models and could add logging tomorrow, but they carry the same conflict one level up: they sell all the vendors, so they cannot publish the cartography of which ones fail where. And their logs live inside the platform whose lock-in is the business model. A record you would stake a dispute on has to be owned by the customer and exportable, which is the structural opposite of a platform log.
Only an independent party can sit in this seat. Only a non-airline could have built the reservation system every airline trusted.
The layer above is obligated. The layer below is conflicted. The middle must be independent.
Two ways in, one substrate. Run the session inside Spegling, in a governed space it operates, reachable from any browser and still working after you close the laptop. Or keep the AI you already use, Claude, ChatGPT, a coding agent, and wire it in over MCP. Either way the work goes through Spegling, and Spegling keeps one memory across all of it, so tomorrow's session knows what today's was about.
Email is the sharpest example. Your mailbox connects to Spegling once. When your Claude session goes through the inbox, it holds no mailbox credentials; it asks Spegling's gate. The key it carries is bound to one role, so the founder's assistant reads the founder's mail and nothing else. Reading and drafting pass. Sending stops at the gate and waits for you. Every access lands on the Chain, so a month later you can answer exactly which messages any AI read and what it drafted.
Every tool follows that shape: connect once behind the gate, let reversible work run free, stop what cannot be taken back, review the important work with a model from a different family, and seal every step to the Chain.
The work
Sessions running inside Spegling, or your own Claude, ChatGPT and coding agents wired in over MCP.
Spegling
One memory · the gate · different-family review · the Chain you own.
Your world
Mail, calendar, files, tools. Connected once. Credentials never handed to the caller.
Version control is software's system of record, and it answers one question: what changed. That was enough while the author, the committer, and the reviewer were the same person or their peers. Agents ended that. They already write a large share of new code, the share grows monthly, and git has no field for the questions that now decide trust: which model wrote this, against what stated intent, who verified it and from which family, at what cost, and whether the thing reviewed was the thing that shipped.
Spegling binds those answers to git instead of replacing it. Every agent-written change carries a pointer to its sealed row; every row carries the commit, the reviewed tree hash, the verdicts, and the cost. Blame on any line ends in the record. Merging is the irreversible act, so that is where the gate sits.
History says what this becomes. When trade outgrew personal witness, double-entry bookkeeping arrived and the audit profession grew around it. When flight outgrew the pilot's account, the flight recorder became law. Software production is outgrowing human witness now, and its second ledger is missing. Development is the first market because the artifacts are already hashable and the buyers are the builders. The same primitive then follows agents into every workflow they take over: mail, finance operations, legal drafting, support.
A side effect that earns its own slide: cost per feature. The ledger knows what each shipped feature cost in tokens, euros, and human minutes. That number exists nowhere today.
The EU AI Act requires high-risk AI systems to keep a continuous, tamper-evident, traceable log across the system's lifetime (Article 12). That is a legally required ledger, phased in on a known clock.
Public enforcement runs behind the clock. The first obligations bite in 2026, while the EU's third-party evaluation capacity is expected operational only in 2027 and most member states were late designating their authorities. Between obligation and public verification sits a window, with a named end date, that private evidence tooling fills.
2 Aug 2026
Article 50 transparency: tell people they are dealing with AI, label generated content.
2 Dec 2027
High-risk obligations, Annex III standalone systems. Record-keeping becomes law.
Aug 2028
High-risk obligations for AI embedded in regulated products.
The window is time to build the record in. Spegling produces the record and the review that compliance work stands on; the certificate is someone else's to issue. The architecture is EU-resident, built in Finland.
Demand for the record runs on three clocks. Trust is one: organisations scale agents only as fast as the verification loop earns it. Law is the second: the timeline above. The third is money, and it arrives on market logic rather than a compliance calendar.
Payment rails for agent commerce went live ahead of the economy itself. The x402 protocol had processed 165 million agent transactions by April 2026; the Agent Payments Protocol now sits with the FIDO Alliance, backed by the major card networks. The volumes are still small, about $50 million cumulative. The direction is set, and the protocols share a blind spot: signed mandates prove what an agent was allowed to buy, the rails move the money, and nothing in the stack proves the delivered work was what was paid for.
When real money flows through agent transactions, disputes follow, and disputes run on evidence captured at the moment of the act. That is the layer Spegling already builds for human oversight, and the same sealed record is what lets two companies' agents transact: outcome-priced, evidence-settled. Payments stay someone else's business. The Chain's evidence stays exportable into whatever settlement stack wins, a format decision made now for a market that arrives on its own schedule.
This page's first reader is increasingly an agent doing due diligence for a buyer. We like our odds with a reader that only counts evidence.
Every governed run writes a labeled row: this maker, this task type, caught or missed by this reviewer family, at this cost. No vendor can assemble that table, because it spans vendors. It compounds:
Rows
The ledger accumulates verdicts across models, families, and task types.
Cartography
The failure map of the model market: who fails where, caught by whom.
Routing
Learned reviewer routing: the right checker for this work, priced by evidence.
Underwriting
Attestation of agent output, and early warning when a model regresses.
runner → intelligence → underwriter, on one accumulating table. The moats are functions of row count.
Copying the plans does not copy the position. The questions a reviewer asks are crystallized judgment, accumulated per task type and industry the way actuarial tables are: the form is public, the tables are the company. Acceptance by auditors and insurers is earned one relationship at a time. And a ledger's age cannot be replicated by anyone, including us: a record kept since 2026 cannot be backdated, so every month of operation widens a gap money cannot close. A well-funded copy starts at row zero on all three.
The table serves three buyers, in order. Today, the builder who hands real work to agents and keeps the wheel. As the obligations land, the organisation that must answer for what its agents did. Then the widest ring: organisations with no agents at all, most of the economy. They have been waiting on exactly what this layer supplies, and they arrive with an advantage the early movers spent: nothing to retrofit. For them Spegling is the whole stack, sessions to roster to gate to record, every person under their own hat, the org owning its Chain. Velocity is the wedge; obligation is the expansion; the governed start is the volume.
A v0.1 governed loop in production on a private allowlist: maker proposes, a reviewer from a different family answers fixed yes/no questions, the row seals onto a hash-linked, exportable Chain. Mail and calendar connect behind a gate that never hands over credentials. Consulting revenue funds the substrate, and design partners are being recruited from regulated EU organisations.
Distribution is the engagement itself: each consulting engagement discovers a client's real workflow, ships it as a governed agent, and seeds a tenant whose ledger starts accumulating from the first run. The retainer that keeps the agent honest is the platform subscription. Services revenue and platform adoption are the same motion.
A record written in the moment is evidence. The same story reconstructed after a dispute is testimony.
Named comparables and their valuations live in the dated deck. They move; this page stays.
Hannu Varjoranta. Systems engineer, founder, writer. Two decades of infrastructure, security, and data systems, including Spotify and F-Secure; co-founder of Valo and Cloop; model-compression work merged upstream into vLLM.
The multiplication is the product itself. The process that builds Spegling runs as a governed loop with the founder as principal: deploy authority sits behind gates, feature status flips on deterministic probes, and the loop is now pointed at Spegling's own repository. One person, one honest queue, and an audit trail for the code that builds the product.
Varjosoft is deliberately small and builds patiently. Consulting revenue funds the substrate, and the ledger grows ahead of the story. If there is ever a round, it prices evidence.
The essays are written with a pen first. The site you are on runs on the conviction it sells.
No process, no deck. Write a line about how you see the layer: hannu@varjosoft.com. A person reads it and a person answers.