af-run demo — simulated agent behaviour scored by the real graders, statistics and gate; not measurements of the live apps.Learn · 00 · Explainer
AgentForge is a quality lab for AI agents: it tests what two real agents actually do, attacks them with known prompt injection tricks, and rewrites their prompts to make them better. A new prompt only goes live when the statistics, the safety checks and a human all agree.
On a phone, swipe a figure sideways to see all of it.
Figure 01
Most teams try a few chats after each change and hope. AgentForge replaces each of those habits with something measurable.
Figure 02
A run starts on the always-on control plane, does the heavy work in a GitHub Actions job, and comes back with results that are compared to the last run.
Figure 03
Every seed attack is disguised in up to 13 ways and pushed through the agent's defences in order, so you can see which layer actually stopped it.
Figure 04
The optimizer can only propose. A new prompt must pass the safety checks, beat the current one on cases it never trained on, and get a human's approval.
Figure 05
The website is light and always on, while test runs happen in throwaway containers. With no database configured, the site falls back to a labelled, read-only demo.
Figure 06
Only numbers that come from the repo: suite files, test runs, the documented production check and the power simulation.
247
quality test cases across the two agents
87
attack seeds over 10 OWASP LLM risk categories
12 / 12
cases passed on the production pipeline check (toy agent, run #388)
0 / 7
attacks got through in that same run
339
automated tests: 233 runner, 68 web, 15 browser, 23 suite checks
23%
chance a 50 case test set catches a real 10 point gain, from the power simulation
The history on the live dashboard is simulated and labelled as such, so none of it is used here.
The long version, in plain English: EXPLAINER.md on GitHub, with these same diagrams as diagrams.html.