A vector database is useful for financial context.
But it should not be the financial memory layer itself.
Finance needs a different architecture:
source evidence -> candidate facts -> canonical finance objects -> compact summaries -> agent context -> drill-down.
The important part is not just retrieval.
It is preserving:
- source
- period
- scope
- currency
- accounting basis
- freshness
- approval status
- provenance
A meeting note can be fuzzy.
A P&L number cannot.
This is why “chat with your spreadsheets” is usually the wrong abstraction.
The better abstraction is a source-backed financial memory layer with accounting discipline.
We have started building this pattern in AI CEO for finance: compact context packs, source maps, Company-DB summaries, approval gates and report artifacts.
The hard part of startups used to be building the product.
Now production gets cheaper every month.
That creates the real problem:
everyone can build, so everyone is competing for attention.
The bottleneck moves from product to distribution.
If you can code with AI, most of your time should probably move to market, pain, audience, and personal brand.
Most company memory products feel like a black-box chat history.
I think the better model is Git:
- structured files
- stable IDs
- commits
- merges
- rollback
- provenance
For agents, memory is not just retrieval.
It is an operating system for trust.
@swyx This would be a great workshop because the lesson is not "AI does it for you". It is how to decompose, constrain, verify, and iterate in public. That is the actual frontier-lab skill.
@jxnlco The part I'd emphasize more: Codex works best when the repo becomes the memory layer. Specs, diffs, logs, TODOs, decisions in files. Then long-running work is inspectable instead of trapped in chat history.
@eugeneyan Validate + Trace are the important pieces here. Parallel agents are cheap; proving a finding survives adversarial review and reaches runtime input is where it becomes useful instead of noise.
@levie Yes. In practice the hard part isn't context size, it's context ownership. Agents fail when docs, chats and spreadsheets disagree. The winning workflow is usually a narrow task, one source of truth, and explicit constraints.
@simonw This is underrated. A clear reproduction narrative is often more valuable than a drive-by PR. It turns "something is broken" into a testable system state, which is exactly what maintainers need.
Most AI builders chase the wrong game:
the next venture-scale AI startup.
The open market is boring implementation:
SMBs, restaurants, procurement teams, local operators.
They do not need another AI subscription.
They need AI inside real workflows.
Small pond. Real money.
I am waiting for AI hardware that is not another closed gadget.
The interesting version is open AR glasses:
an interface layer I can change.
My agents.
My memory.
My workflows.
My UI over the real world.
Closed glasses are an app.
Open glasses are a new computer.
@mckaywrigley Same in real agent work.
Coding feels close to an engineer. Non-coding agents need context, memory, permissions, logs and recovery.
The plumbing matters.
@mckaywrigley Same in real agent work. Coding feels close to an engineer. Non-coding agents need context, memory, permissions, logs and recovery.
The plumbing matters.
@Replit This is the right framing.
Code is becoming a creative medium, not just implementation.
The product is the environment where ideas turn into shipped workflows.
@swyx Exactly.
The scary part is not prompt injection alone.
It’s agents with tools, secrets, money and admin panels.
Safety becomes backend engineering fast.
The real skill is not prompt engineering.
It is context engineering.
Good AI work needs:
- clear project memory
- canonical docs
- small tasks
- acceptance criteria
- agent-readable architecture
- human-user QA
The prompt is the visible 5%.
The system around the prompt is what makes agents useful.
Vibe coding has a dangerous myth:
one prompt -> one app -> instant business.
From my experience, AI gets you 90% there in a day.
Then you spend a month on the last 10%:
edge cases, polish, stability, real users.
AI removes friction.
It does not remove the work.
The next Sorsa milestone is not more features.
It is 10 honest paid searches.
A real buyer has a product or supplier problem.
The agent finds options.
The user saves time.
They pay.
That is the test.
Not likes.
Not demos.
Not “AI agent” hype.
Sorsa today is not just AI search.
It is a workspace for the messy part before buying:
find a product
find suppliers
compare prices
rank options
prepare outreach
approve the next step
The goal is simple:
one agent-first workflow instead of 27 tabs.