Agents you can be accountable for.
PAC is an open framework for governing agentic AI: three pillars,Potential, Accountability, Control, six profiler axes, and nineteen questions your team should be able to answer with evidence rather than policy documents.
One conviction runs through all of it: autonomy is not a setting, it is a track record. What an agent may decide on its own is exactly what it has proven, and the bar rises with how far the damage reaches when it is wrong.
authored and maintained by AboveBeyond, at home ontrustedagentic.ai, run daily in production byElixir
Three pillars, each one question.
Potential
What’s worth building that lasts? Find where agents create real value, and what over-constraining them is quietly costing you.
Accountability
Who’s accountable, and can you prove it? Every agent known, every consequential decision traceable to who authorised it.
Control
Can your infrastructure enforce what policy demands? Contained by architecture, not by rules an agent can ignore.
Six axes, nineteen questions.
Every agent, playbook or automation gets profiled on six axes, from how autonomous it is to how far the damage reaches when it is wrong, and every level assigned needs evidence. Then nineteen questions your team should be able to answer from that profile. The full framework, with the scales and the profiler, lives on its own domain so it can be adopted without standing on a vendor’s site.
We profile yours.
Bring us the agents you run, or the plans for them. With your team we profile every one on the six axes, set the controls it has to stay within, and set the ceiling it has earned. You leave with something you can prove, and with the people who run those agents knowing why. Elixir is where the same rules govern a real fleet every day.
an agent fleet to profile? talk to us
Don’t read it. Run it.
The framework ships as a free scan for Claude Code. It inventories everything in your repository that acts on its own: agents, LLM calls, MCP servers, bots, scheduled jobs. It scores each on the six axes with file-and-line evidence, and closes with one verdict line.
/plugin marketplace add abovebeyond-ai/plugins /plugin install agent-scan@abovebeyond /agent-scan
Its sibling /hallucination-scan covers what agents say rather than what they do, on the ProveML method: every claim traced to a fact, every number sourced, every judgment held against a declared threshold.
Signed, not neutral.
PAC is authored and maintained by AboveBeyond. It has its home ontrustedagentic.ai so that others can adopt it on its own terms, and it is signed rather than dressed up as an institution, because frameworks earn adoption through people with receipts. Ours isElixir, which has governed a fleet of client systems under these rules in production, ledger and all.
The parts that need real neutrality will get it the real way: theProveML verification method is published under Apache-2.0, and where it should live long term is a conversation we are having in the open.
questions? talk to us