Published experiments
Copied from Atlas with each recap's local assets and source fragments.
2026-11-10
PR 3801 after start_reply: what the person sees now
The rerun on top of the new reply mark, drawn as the Studio chat: the hidden note, the forms that pass, and the forms with no sentence that do not say what they are about.
2026-11-09
PR 3801: the QA failures, as a person sees them
Eleven failed checks drawn as the Studio chat shows them, with the real words: what is wrong, whether Studio shows too much or the agent wrote it, what the judge read, and a one-line fix.
2026-10-23
Better research. Better pages.
Five improvements and two new routines for Fimo’s SEO, GEO and Content agents, with an interactive routine gallery, external data sources, worked example and pilot plan.
2026-10-22
Signal experiments: what 139 production chats taught us
Jev, a small model and a strong model read three days of production chats: cost per chat, what each reader catches, 26 grouped findings, and why plain code over tool errors found the clearest bugs.
2026-10-22
The software factory: Fimo finds its own problems
A team pitch: tests and live checks find problems, coding agents open the fix as a PR, and people review and merge. With a real before and after, and the path from monitoring to safe automatic merges.
2026-10-20
Admin assistant: saved threads and rich blocks
PR 3214 for Fimo Admin staff: saved threads on the chat engine, History, 12 block types shown in real screenshots, single-use confirmation cards, the isolation and cost guards with their security review, four live QA passes, and an honest list of what the agent does well and what it misses.
2026-10-19
Fimo agent templates: the parts and how they connect
PR 3170 architecture: the six blocks of an agent (template, Studio skill, shared crawl, turn, data tools, gallery), their files, JSON shapes and API contracts, why a template copied once into the database never updates, and a target shape with one folder per agent and instructions read live.
2026-10-16
SEO ideas inside Fimo's agents and routines
Follow-up to the 8 ideas, on today's develop: the SEO, GEO, Content and Quality agents, their memory, the 7 routines waiting for data, and where AI answers, competitors, page 2 and fact checks fit as routines.
2026-10-16
SEO agents: 8 ideas to make them better
A proposal for beginners in SEO: fact checks, fixes for wrong AI answers, agent memory, and page 2 to page 1. Each idea has an example and a diagram, plus a 3-phase plan.
2026-10-15
Outrank teardown
How Outrank grows a site from outside: daily AI articles, a backlink credit network and a Search Console refresh loop. Screen by screen, then what Fimo SEO and GEO agents do differently.
2026-10-15
Agents that remember
Tell Fimo a rule once and every chat follows it. Three levels of memory, how it works inside, the tool and why, and how we tested it. Step 1 reads in 5 minutes.
2026-10-13
Needs you: the chat keeps track of what matters
Questions from many routine runs get lost. Proposal: one short list the agent keeps current, with Studio prototypes.
2026-10-12
Agents that wake up on their own
Smart triggers, the website watcher, the agents we build next, how it scales, and how we iterate on v1. Step 1 reads in 10 minutes.
2026-10-09
How Fimo events start routine runs
Who listens, what triggers a routine, how the engine runs it, and how API replicas share the work.
2026-09-30
Agent reviews: why they break, and the fix
Why autonomous-agent review requests hide their changes, go unnoticed, and leave branches open after a merge, with one proposed fix per gap.
2026-09-04
How Studio chat survives a deploy
The high-level code architecture from SIGTERM and safe model-step settlement to durable continuation, browser reconnect, and one continuous visual reply.
2026-09-03
Agent work across deploys
The full timing and recovery story for Studio chat, Redis replay, the 13-minute shutdown budget, and autonomous Branch continuation.
2026-09-03
Astro and SvelteKit SEO settings
Updated September 5: pure SEO conversion, app-owned metadata, framework recipes, agent guidance, and generated-output proof.
2026-09-03
Website quality for Fimo agents
How agent-browser, Lighthouse, and production evidence could fit together without confusing lab and field signals.
2026-09-02
Progressive block experiment
One recap block reveals more context and proof in place while preserving the continuous review story.