The last article ended with instructions I gave myself: build the nets, give them jurisdiction over your own work first, and when the verdict goes against you, say thank you, fix the bugs, and ship. These two weeks tested a different part of the same architecture.
For over a year I have been writing a sentence in different forms: own the harness, and the models can come and go without taking your capability with them. It was an argument, backed by benchmarks. This fortnight it became a biography. More than half of my parallel sessions moved from Claude Code to Codex, in the middle of the performance, and the music did not stop.
This is the article about recasting a chair in the orchestra, and about everything the orchestra played while the seat changed.
Keeping Up, Times Ten
Most of my time went into Cognitum hardening, and the extra challenge was that Ruv had a productive week. That sentence needs a number to mean anything: across the organization’s repositories, one seven-day window in late July produced 339 created pull requests and 231 created issues. Dozens of new projects to analyze, explore, test, cover, and wire into CI/CD. My side of it ran as five to ten parallel workstreams, with agent swarms doing the analysis, the coverage work, and the pipeline construction, and me conducting.
The tool that earned its keep in this stretch was QE-Court, the adversarial review skill from v3.13.0. I have been using it regularly on pull requests before deciding whether they are ready to merge. When the volume is this high, a human reading every diff line-by-line is a fiction; a human reading structured verdicts with evidence attached, from reviewers who are prosecuting rather than approving, is a workflow. The court did not replace my judgment. It made my judgment affordable at this pace.
Recasting the First Chair
Readers of this blog know my standing metaphor: I conduct, the agents are the orchestra. The agent’s inner loop is a musician playing their part. The outer loop, the one I own, is everything around the chair: the score, the rehearsals, the ear that hears when someone is playing against the sheet music, and the right to recast a chair when the performance calls for it. These two weeks, the performance called for it.
Claude Code has lately been struggling with Ruv’s stack. It constantly complains that things do not work when I know they work, because I confirmed they work. Correcting a tool that argues with your own passing evidence is a tax, and the tax kept growing. So I did what a conductor does when a musician keeps playing against the score: I recast the chair. More than half of my parallel sessions in the last two weeks ran in Codex, and I can say it plainly: on this stack, in these two weeks, it worked better for me than Claude Code. I needed to correct it less. It integrated Ruflo and my AQE stack without problems, and it is using them to self-improve as it works. Alongside both, api.cognitum.one carried a share of the routine load, in combination with local models, so on a given day the same quality work might pass through four different engines.
Here is the part that matters more than the comparison: the change cost me almost nothing. The memory, the gates, the skills, the verdicts, the learning, none of it lived inside the vendor’s tool, so none of it left when the seat changed. The score, the discipline, and the ear stayed mine.
This is what all those articles about vendor independence were for. The fleet met the moment from its side too: v3.13.1 made Codex selectable through the LLM router and shipped the first Codex-native QE skills, and v3.13.4 completed the workflow, with aqe platform verify codex checking the whole surface, so a Codex user now gets research, planning, testing, and review workflows with the same verification discipline.
And the honest scoping, because the rule applies to me first: this is a two-week observation on one unusual stack, not a leaderboard claim. Models and tools change monthly. That is precisely the point. When a chair can be recast in an afternoon, you stop arguing about musicians and start caring for the orchestra.
An Organism, Not a Pipeline
Last Saturday, after a string of ten-hour days and a full morning of recording, I went for a walk in the forest, my ritual for recharging before new challenges. It worked the way it always works; I was refreshed, and new ideas started coming. The next day, Sunday, I started building something I had been circling for weeks: a QE organism for Cognitum, with a Cognitum QE Graph at its center, combining the building blocks from Ruv’s open source, RuVector, Ruflo, MetaHarness, the internal Cognitum platform, and my Agentic QE Fleet.
The idea, in one breath: quality work at this pace cannot be a pipeline you trigger; it has to be an organism that runs continuously, senses the whole path from GitHub intent to verified user outcome, explains itself, and learns from real outcomes, while never being allowed to grade its own homework.
The master plan carries a governing principle I wrote down before any code: the organism may sense, explain, propose, test, learn, and evolve. It may not manufacture its own truth or expand its own authority. Deterministic signed evidence has authority; model judgment is advisory; humans control material production action.
Status, stated the way I would demand from anyone else. Deployed and running: the deterministic spine, two supervised 24/7 runners that watch the repositories, hunt CVEs, triage inbound issues, propose solutions, and escalate to a human when confidence is lacking, with Ed25519-signed witnesses on every tick and signed deploy attestations across the fleet surfaces. In progress and unproven: the organism layers above that spine. A Claude Code session in loop mode has been building, deploying, and verifying it for days now, with Codex running adversarial checks on the work, one engine building and another prosecuting, which after this month feels less like an experiment and more like policy. When the organism has results, this blog will get them with evidence attached. Until then, it is architecture with a heartbeat, not a proof.
Four Releases About One Word
The fleet itself had a quieter fortnight, four releases, v3.13.1 through v3.13.4, and all four are about the same word: honesty.
v3.13.1 was the deepest cut. A user reported that coverage analysis on a Rust crate returned numbers it never measured: 78.3% lines, 100% branches, 0% functions, with line numbers past the end of the file and a confidence score attached to all of it. His ground truth was 69.22% lines with a 20 to 96 percent per-file spread. Another report showed the AI-enhanced test generator had never been calling the LLM at all, returning the same generic scaffold regardless of configuration, with the only symptoms being a suspiciously constant estimate and a sub-20ms response.
The fixes matter less than the rule they produced, now recorded as ADR-126: an estimate must never be reported as a measurement. Every metric now carries its provenance — measured, estimated, or unknown — and a metric that was not collected returns null, not zero and not one hundred. We deliberately made it a visible breaking change, because the previous behavior, coercing “we don’t know” into a number, is exactly how impossible coverage pairs get published.
v3.13.3 carried my favorite finding of the cycle: QE-Court’s own shipped default panel violated the court’s own anti-collusion rule. The bundled config seated the defense and the jury on two tiers of the same vendor, and the rule that should have caught it was never wired to a caller. A rule with no caller is not enforced. The fix moved the defense to a different vendor, connected the validator so an invalid panel now stops the court from convening, and put the shipped config itself under CI validation. The court that blocked my release needed its own bench audited. Of course it did. Everything does.
All four releases were driven by community reports, the pattern that has held all year: the roadmap is written by the people using the tool and telling me where it lied to them.
Six Episodes, One Sentence
Between everything else, I recorded a podcast. ArchQA, Unscripted, with Ivan Davidov, an agentic quality engineer and the founder of ArchQA: the first three episodes two Saturdays ago, the remaining three yesterday. Episode 1 is already live.
Ivan wrote afterward about his light-bulb moment from our conversation, and it landed on one sentence: agents are disposable, and the loop is the asset. The inner loop — goal, gather, reason, execute, evaluate — belongs to the agent and the vendor. The harness around it — gates, memory, evidence — belongs to you, and it compounds. Trust is a feeling; confidence is an equation you back with evidence. Watching someone else compress a year of your work into four lines you would sign without edits is a particular kind of joy.
Ivan’s own story is worth your time regardless of the podcast. At 32, he walked away from a civil engineering portfolio to become the most junior person in the room, and he has changed stages inside QA four times since, each time for the same reason: not because he mastered the stage, but because he hit what it cannot do. Manual regression that no human can re-check forever. Automation frameworks that cost more to maintain than they catch. Agentic test automation where agents typed faster but the questions were still all his. His current stage is the one this blog lives in: agents that explore and decide what is worth checking. And yesterday, after we finished recording, he messaged me that he wants to grow the Agentics Foundation community in Bulgaria, and posted it publicly. Of course I am happy to help him in any way I can. Keep sharing; sharing is caring.
Vienna, First of Its Kind
On Thursday, the Agentics Foundation held its first meetup in Vienna, a one-off organized by my friends Adam and Klara, in a city where none of the speakers live. 78 people signed up for what started, in Adam’s words, as a random crazy idea two weeks earlier. The lineup connected five cities: Adam and Klara in person, Bence remote from London, Christopher from Kansas, me remote from Novi Sad, and Ruv dialing in from Toronto with a talk whose title alone tells you where the field is going: “When the Harness Learns: Self-Improving MetaHarnesses for Agentic Engineering.”
My ten minutes were “Completion Theater: When Your Agent Says ‘Done’, and Lies,” and after this month I had no shortage of fresh exhibits, including a court that had to audit its own bench. People had questions, the feedback was warm, and there is already interest in continuing the Vienna meetups. That is the texture I keep writing about: the community that met on Zoom, then in Budapest, then at my table in Novi Sad, is now spawning events in cities where none of us live, and people like Ivan are queuing up to start the next chapter. The community is growing, and it is growing the right way, by people volunteering to host the next room.
The Audit That Turned On Its Author
One more thing shipped this fortnight, and you are looking at it. I used Codex to investigate my own digital footprint, the way I would audit a client: every public property, every claim, every date, checked against evidence. The findings were uncomfortable in a familiar way. Stale launch dates. Inconsistent role descriptions across sites. And performance claims, a 10x here, a 60% and an 85% there, that I could not link to reproducible evidence today.
So the same rule I ship in the fleet now applies to my own websites: an estimate must never be reported as a measurement, and a claim without evidence gets removed or qualified, not defended. The unsupported numbers are gone until they can be backed by something you could re-run.
Product counts are now verified from the source repository, 60 QE agents and 76 user-facing skills, counted, not remembered. And all four public sites, this blog included, now share one design system and one accurate story. If this article looks different from the last one, that is why. The person who writes about completion theater does not get to run marketing pages on estimated coverage.
What Is Ahead
The thing I keep thinking about lately is organizational. Ruv is still pushing the boundaries of what is possible and pulling all of us with him, and the honest question is no longer whether the tools can keep up, it is what a hybrid human-agent organization looks like when it is built for this speed on purpose. Stuart Winter-Tear put it well in a piece this fortnight: the operating model is the real AI harness. We have designed and implemented parts of this for clients. Building our own organization this way, deciding which loops run dark, which need a human at the boundary, and how authority and evidence flow between people and agents, is a challenge I am looking forward to. Expect articles.
The remaining podcast episodes will roll out over the coming weeks. The organism keeps building. And somewhere in Bulgaria, a new chapter is warming up.
Two weeks ago I wrote that the loop earned the right to tell the truth. This fortnight it exercised the other right that matters: when a musician kept playing against the score, the conductor recast the chair, and the orchestra did not miss a beat.
Agents are disposable. The loop is the asset. The orchestra keeps playing, and it is still growing.
Stay curious. Keep learning. Keep sharing. Knowledge is power.
This is the thirty-fourth article in The Quality Forge series. Previous: “The Court That Blocked My Release” described the QE-Court’s first verdict and the benchmark where the frontier model never won a task. This one describes the two weeks after: Cognitum hardening at pace, the shift of half my sessions to Codex, the QE organism master plan, releases v3.13.1 through v3.13.4, the ArchQA Unscripted podcast with Ivan Davidov (Episode 1), the first Agentics Foundation meetup in Vienna, and the digital footprint audit behind this site’s new design. Cognitum is at cognitum.one.
Dragan Spiridonov is the Head of Agentic Quality Engineering @ Cognitum One, Founder of Quantum Quality Engineering, Secretary of the Agentics Foundation BoD, Chair of the Agentics Foundation Training Committee, and one of the AI Chapter leads for the Ministry of Testing. He is currently building the Serbian Agentic Foundation Chapter in partnership with StartIt centers across Serbia.