AKB demo access

Enter the demo passcode to open the portal. This gate is separate from your Cursor API token (you paste that later for live skill optimization).

Agent Knowledge Base

Portal

This page turns recorded agent sessions into grouped knowledge, then into a rewritten skill for the next model.

Do these in order
Featured example
Episode — one agent session (the rows below).
Cell — a bucket of similar sessions (who × why × topic).
Failure mode — a repeated way of failing, with a count.
Skill — the instruction file we rewrite for the next model.
Local v0 · no GPU · no API key needed
Software specification · storage, APIs
Older synthetic process demo
Step 1 — Browse the recorded sessions

The four numbers are buttons. Click 29 failed to see only broken sessions. Click a row to read the conversation on the right. When you are ready, press the gold button — that is the only action that builds knowledge.

Collected traces

Recorded agent sessions from several products, stored as episodes.

Compresses every session into cells and lists failure patterns. About one second. Does not call an LLM.

Where the data came from

Click a source to show only its episodes. Click the number 6 sources above to come back here.

What happened in this session

Empty until you click a row in the table. Prefer a red fail row — those feed failure modes.

Click an episode row in the table below.

Every session (episode)

Click a row to open it. Use the chips or the numbers above to filter.

Step 2 — Read the grouped knowledge

A cell is not a spreadsheet cell. It is a bucket: same kind of user, same kind of task, same topic. Click failure modes to jump to step 3, or press Next.

AKB knowledge

Grouped sessions. Dual representation: example traces + structured failure modes.

This example focuses on failures. Cells with 0 failures are the success contrast.
Compression = tokens dropped by collapsing repeated prefixes and keeping representative traces. Real collected traces compress less than the synthetic demo because the tasks are more diverse. Click the number again to hide this.

Cells — click a row that has failures

Opens the failure-mode list filtered to that cell. Accept is the share of sessions in the cell that succeeded.

Step 3 — Pick a repeated failure

Support is how many failed sessions look like this. Click a card (or the gold strip on it). That does not run anything yet — it only opens the skill-optimization form with this failure filled in.

Failure modes

Repeated ways the agent broke. This portal’s working example.

Filtered to one cell.
Featured: create-pptx skill optimization for Composer 2.5

Bundled failure: wrong slide layout index on the title slide. Without a Cursor key the server runs a deterministic mock teacher so you can verify iteration diffs locally.

Uses student composer-2.5 by default.
Step 4 — Rewrite the skill for the next model

Same as eval-platform’s judge token: paste or upload a Cursor API key, click Test connection, then pick teacher / student / optimizer from that key’s catalog. Each iteration writes a new SKILL.md; the result panel diffs it against the previous file. The key is this-run only — never stored in the job JSON.

Skill optimization

Start from Failure modes if this form is empty.

Disabled until you click a failure mode.

Job config — what you are asking for

No failure selected yet. Go to step 3 and click a card.
Cursor token · this run only

Eval-platform rule: never logged, not written to jobs, cost rides on your token. Upload a file or paste. Optional base URL if you go through an OpenAI-compatible gateway.

Not connected. Run optimization will use the local fake teacher.

Diagnoses the failure. After Test connection this list is the models your key can actually call.

Gates the rewrite (no-retry / preservation). Same catalog as teacher once the key is live.

Rewrites SKILL.md. Can be the same model as teacher.

0.2 is conservative. Eval-platform pins temperature for comparable runs.

Each pass diffs new SKILL.md vs the previous file. Eval-platform asks for ≥3 LLM samples; default 3 when a key is connected, 1 for the local teacher.

Result — per-iteration file diff

Each block is one rewrite: diagnosis, student gate, then unified diff of old vs new SKILL.md.

Nothing here yet. Connect a key (or skip it for the local teacher), then press Run optimization.
How this is built: Software specification — what is stored where (no SQL database in v0), HTTP APIs, and the four platform contracts.