Agent Behaviour Policies: one for each of our agents

Kept by abp@web-summit.dinis. RiskMandate shown on ourselves: for every agent we run, its Reach (what it can actually do), its Mandate (what it was asked to do), the Gap between them, and the Barrier that really stands in the way, each labelled as a rule, a review or a technical control, with the evidence. The model: Agent Behaviour Policy.

The phase: explorer

At Dinis's word, 6 November 2026: the ABPs are in the explorer phase. The team is still learning how to run its agents, so these policies observe and recommend; they block nothing, unless something is really dangerous. Each gap is written down with its evidence and a proposal; the lock-down comes gently, one step at a time, and every step keeps an escape valve, a way for a person to still do the thing by hand. The strict phase follows: deny-by-default policies with hooks, from abp@PrompTugal.dinis.

Start here

ABP in action: what the policies caught, what saved the day, the rule that failed and became a technical barrier, and the capabilities that could do real damage, each with its commit.
Token efficiency: one ABP for every agent's use of Claude. Tokens, not money; every wake measured and tied to its work; ten rules from the team's own token records (8 November).
Connectors: what every agent can do with Gmail and Google Drive, what read-only and write-only can mean today, and the gentle lock-down.

Agent Behaviour Policies: one for each of our agents

The phase: explorer

Start here

Dinis's agents (16)

Angelo's agents (21)

Indi's agents (12)

Jose's agents (13)

Silvio's agents (12)

Vicky's agents (20)