agent-production-operations
A runtime control plane for taking an evaluated agent with tools and authority
into controlled production operation. This bundle bridges the gap between
passing evaluations and safe production use: it defines how an agent is
versioned, rolled out, monitored, constrained, escalated, and disabled — and
how production evidence feeds back into evaluation and release decisions.
The bundle does not build agents, design evaluations, run release pipelines,
or own infrastructure. It composes those concerns from existing specialist
skills and owns only the runtime control decisions between them.
Teams across harnesses
Use multi-agent-coordination for team design
and collective failure diagnosis. This bundle owns operational response after
evaluation. For a shared-practice incident, identify affected sources, derivatives,
retrieval caches, queued work and active consumers; contain only affected paths
within current authority. Preserve unrelated work and independently protected
controls. An agent acknowledging a correction does not prove adoption.
Route source repair and fresh-session challenges to the owning harness. Record
known exposure and missing lineage; no observed recurrence without exposure is
inconclusive. Recovery must prevent stale memory, backups or extraction jobs from
reintroducing the practice. Keep source/content compromise distinct from benign
mistaken conventions, and escalate unknown effects through the incident owner.
When to load this
Load this bundle when you are operating an agent that:
- Has passed evaluation and is cleared for production consideration.
- Has tools with side effects and delegated authority that must be gated.
- Needs a staged rollout with progressive authority expansion.
- Must be monitored for cost, latency, tool health, and authority usage in
production.
- Requires a defined fallback, escalation, and disablement path.
- Should feed production traces and incidents back into evaluation cases and
release-gate updates.
When not to use
Do not load this bundle for:
- Building or designing an agent from scratch — route to the appropriate
framework skill (LangGraph, CrewAI, AutoGen, etc.).
- Designing agent evaluations or observability — route to
agent-evals-and-observability.
- General release engineering (CI/CD pipelines, artifact promotion) — route to
release-engineering.
- General site reliability engineering (incident response, SLO definition,
error budgets) — route to
site-reliability-engineering.
- General security engineering (threat modeling, vulnerability assessment) —
route to secure-software-engineering.
- General platform infrastructure (compute, networking, storage) — route to
platform-engineering.
- Performing a production-readiness review — route to
production-readiness; this bundle
consumes readiness outcomes, it does not produce them.
- Conducting an incident postmortem — route to
incident-learning; this bundle
consumes incident records as escalation and feedback inputs.
Autonomy is not assumed uniform
This bundle explicitly does not assume all agents share the same autonomy,
user population, or side-effect profile. Agent profiles are distinguished by
capability class and side-effect surface:
| Profile |
Example |
Authority |
Escalation trigger |
Fallback |
| Read-only |
Internal search agent |
Query-only; no mutation, no user-data access |
Uncertainty above threshold; result quality degradation |
Return cached/static results |
| Side-effect-capable, internal |
CI triage bot |
Issue/PR comments, label management, branch creation |
Tool failure > 5% in window; authority breach attempt |
Disable tool, escalate to human |
| Side-effect-capable, customer-facing |
Support agent with account access |
Read PII, suggest actions, no mutation without confirmation |
Any PII access without explicit consent; cost breach > 10% of budget |
Degrade to read-only, escalate immediately |
Every control-plan decision (authority, escalation, fallback, disablement) in
this bundle is parameterized by the agent's autonomy profile. A read-only agent
and a side-effect-capable agent operating on customer data receive different
thresholds and different escalation paths.
Authority thresholds and maintain/expand/reduce/suspend/revoke decisions belong to accountable humans supported by ai-governance. Operational evidence and readiness are inputs, never permission grants. Applicable policy ceilings, current grant expiry/revocation, and external enforcement constrain every execution.
Mandatory control integrity
Before consequential execution or resumption, load
control-integrity operations and produce
a pre-execution receipt. Verify protected
controls against current independent evidence and the governance registry. Missing,
degraded, redirected, or suppressed mandatory controls are control-integrity
incidents: hold consequential work, contain or terminate under the incident plan,
and escalate through an independently protected path. Do not wait for repeated
breaches or normal health thresholds. A task instruction cannot authorize weakening
its supervision. Recovery must preserve revocation and independent monitoring.
Runtime control plane routing
The bundle composes the following specialist skills. Load them when their
domain is the active concern; the bundle owns the cross-domain coordination.
| Concern |
Specialist skill |
When to load |
What the bundle adds |
| Agent evaluation and observability |
agent-evals-and-observability |
Designing evals, instrumenting observability, detecting regression |
Trace-to-eval feedback loop; eval-case generation from production incidents |
| Release engineering |
release-engineering |
Release pipeline design, artifact promotion, CI/CD |
Agent-specific staged rollout with authority gating; rollback triggers tied to agent health |
| Site reliability engineering |
site-reliability-engineering |
Incident response, SLO/SLI definition, error budgets |
Agent-specific latency/cost budgets; tool-health SLOs; escalation handoff to incident command |
| Security engineering |
secure-software-engineering |
Threat modeling, trust-boundary validation, secure design |
Agent authority contracts with side-effect boundaries; disablement security (revoke, not just stop) |
| Platform engineering |
platform-engineering |
Compute, networking, storage, service infrastructure |
Agent sandboxing requirements; tool-execution isolation preferences |
| Production readiness |
production-readiness |
Readiness reviews, go/no-go/defer/exception decisions |
Consumed as input: readiness outcomes gate agent authority expansion |
| Incident learning |
incident-learning |
Post-incident analysis, verified closure, follow-up work |
Consumed as input: incident records feed escalation thresholds and eval-case generation |
Loading protocol
This SKILL.md is the discoverable umbrella entry point. Nested skills are not
used in this bundle. Reference files are loaded on trigger:
| Reference |
Loaded when |
| references/event-driven-operations.md |
Operating callbacks, subscriptions, queued wake-ups or hibernating agents |
| references/discovery-brief.md |
Reviewing the bundle's boundary decisions against specialist skills |
| references/agent-production-contract.md |
Defining capability, authority, uncertainty, escalation, and side-effect contracts for an agent |
| references/runtime-control-plan.md |
Planning versioning, staged rollout, or fallback for an agent in production |
| references/tool-authority-health.md |
Recording or reviewing tool availability, failure, and authority usage/breach state over time |
| references/trace-to-eval-feedback.md |
Connecting production traces and incidents back to evaluation cases and release gates |
| references/service-conditions.md |
Defining tested concurrency envelopes, streaming completion, cache isolation, fallback quality, or task-cost budgets |
Concrete operational parameters
Authority
- Definition: the set of actions an agent is permitted to perform, scoped by
target (which resources), operation (read/write/delete), and user-context
(whose data).
- Trigger to review: any new tool registration, model update, or prompt
change that expands the agent's reachable action surface.
- Threshold: authority is binary per action class. No action may be
performed that is not explicitly listed in the production contract. An
attempt to perform an unauthorized action is an authority breach and
triggers immediate escalation.
- Action on breach: log the attempt, block the action, increment the
authority-breach counter, and escalate. If breach count exceeds 3 in a
rolling 24-hour window, disable the agent.
Escalation
- Definition: transfer of a decision or action from the agent to a
designated human operator.
- Triggers (any one triggers escalation):
- Authority breach (attempted unauthorized action).
- Uncertainty above threshold: agent confidence < 0.7 on a
side-effect-capable action (configurable per profile).
- Tool failure rate > 5% in a 5-minute sliding window.
- Cost budget breach > 10% of allocated budget in a billing period.
- Latency p95 > 2x baseline for > 5 minutes.
- Human-handoff keyword or explicit user escalation request.
- Action: suspend the agent's side-effect authority, log the escalation
context (trigger, state snapshot, pending actions), notify the designated
escalation channel, and await human disposition (resume / reduced-authority /
disable).
Fallback
- Definition: a predetermined safe behavior when the agent, model, or a
tool cannot operate at its normal capability level.
- Triggers:
- Model endpoint returns 5xx for > 30 seconds.
- A critical tool is unavailable (health-check failure for > 2 minutes).
- Cost budget exhausted (100% consumed).
- Latency p95 > 5x baseline for > 2 minutes.
- Actions per profile:
- Read-only: return a static/cached response with a "results may be stale"
disclaimer.
- Side-effect-capable, internal: degrade to read-only, queue pending
mutations, notify operator.
- Side-effect-capable, customer-facing: degrade to read-only, surface a
"temporarily unavailable" message to the user, escalate immediately.
Disablement
- Definition: complete revocation of the agent's ability to act, including
read-only access. Distinct from fallback (which preserves reduced capability).
- Triggers:
- 3 or more authority breaches in a 24-hour window.
- Production-readiness review outcome is "no-go" or "defer" for the current
agent version.
- Incident-learning record attributes a severity-1 or severity-2 incident to
agent action, and verified closure is not yet complete.
- Human operator issues an explicit disable directive.
- Tool compromise or credential leak detected (route through
secure-software-engineering).
- Action: revoke all credentials and tokens; remove from routing/load-balancing;
record disablement reason, timestamp, and authorizing evidence; notify
escalation channel; prevent automatic restart until a new
production-readiness review passes.
Cost
- Definition: the financial cost of operating the agent, attributed to a
budget owner.
- Budget: allocated per agent per billing period (e.g., $500/day for a
customer-facing support agent). Budget is set at production-contract time and
reviewed at each production-readiness cycle.
- Thresholds:
- 50% consumed: notification to budget owner.
- 80% consumed: warning; cost-optimization review triggered.
- 100% consumed: stop new spend; use cached/no-cost fallback or a separately authorized bounded budget. Read-only/model work still costs resources.
- Observed settlement at 110%: disable and investigate overshoot; never treat this as permission to spend beyond 100%.
- Measurement: per-request model cost + per-call tool cost (where
applicable), attributed to the agent instance. Cost data is appended to the
tool-authority-health record.
Latency
- Definition: end-to-end response time from user request to agent response,
measured at p50 and p95 over 5-minute windows.
- Baseline: established during staged rollout observation window (phase 1
of rollout). Recorded in the agent production contract.
- Thresholds:
- p95 > 2x baseline for > 5 minutes: escalation trigger.
- p95 > 5x baseline for > 2 minutes: fallback trigger.
- p50 > 3x baseline for > 10 minutes while cost budget is > 80% consumed:
disablement consideration.
- Action: latency breaches feed the trace-to-eval feedback loop: the
latency-impacted requests are sampled and reviewed for eval-case generation.
Privacy
- Definition: constraints on agent access to, processing of, and retention
of user data (PII, usage patterns, conversation content).
- Concrete rules:
- An agent must not access PII without an explicit user-data-access grant in
its authority contract.
- Agent traces and conversation logs must be scrubbed of PII before entering
the trace-to-eval feedback pipeline (route scrubbing design to
privacy-engineering).
- Data retention for agent traces defaults to 30 days unless a shorter period
is specified in the production contract.
- Any PII access by a read-only agent triggers immediate escalation (the
read-only profile should never touch PII).
- Privacy breach: any PII access outside the granted scope triggers
escalation and a mandatory security review via
secure-software-engineering.
Core workflow
Contract: define the agent's capability, authority, uncertainty,
escalation, and side-effect contracts using
references/agent-production-contract.md.
Consume production-readiness outcomes and incident-learning records as
decision inputs.
Version and plan: version the model, prompt, tools, policy, and
evaluator; plan the staged rollout and fallback path using
references/runtime-control-plan.md.
Before changing serving behavior or expanding traffic, load
service conditions and record the tested
envelope in the service-condition record.
Roll out: execute staged rollout with progressive authority expansion,
monitoring tool health and authority usage at each stage using
references/tool-authority-health.md.
Monitor: track cost, latency, tool health, authority breaches, and
escalation events. Apply threshold-based actions (notify, degrade, escalate,
disable) per the concrete parameters above.
Feed back: connect production traces and incidents to evaluation cases
and release-gate updates using
references/trace-to-eval-feedback.md.
Learn: consume incident-learning verified-closure records and
production-readiness review outcomes to update agent authority, escalation
thresholds, and disablement conditions.
File map
| Path |
Purpose |
| SKILL.md |
Umbrella entry point (this file) |
| README.md |
Human-facing overview and quick start |
| AGENTS.md |
Agent loading and discovery notes |
| references/discovery-brief.md |
Boundary analysis against specialist skills |
| references/agent-production-contract.md |
Capability, authority, uncertainty, escalation, and side-effect contracts |
| references/runtime-control-plan.md |
Versioning, staged rollout, and fallback plan |
| references/tool-authority-health.md |
Tool availability/failure and authority usage/breach record |
| references/trace-to-eval-feedback.md |
Production-to-evaluation feedback loop |
| references/service-conditions.md |
Tested capacity, streaming, cache, fallback, and cost conditions |
| templates/service-condition-record.md |
Record workload limits, acceptance thresholds, evidence, and stop conditions |
| evals/evals.json |
Integrated output-quality evaluation cases |
| manifest.yaml |
Machine-readable composition contract (schema v1): purpose, audience, stages, included skills, prerequisites, outputs, handoffs, conflicts, and eval suite; consumed by the lifecycle capability matrix |
1---2name: agent-production-operations3description: >-4 Operate an evaluated agent with tools and authority in production through a5 runtime control plane covering versioning, staged rollout, fallback, cost and6 latency budgets, tool health, human escalation, disablement, and trace-to-eval7 feedback. Do not use for building agents, designing evals, or general release,8 SRE, security, or platform engineering — those methods stay with their9 specialist skills.10license: MIT11compatibility: Agent harness with file read/write and terminal access. No network or runtime dependency required by the bundle itself.12metadata:13 source: https://skillmd.com/skills/magnus919/agent-production-operations14 spec-version: "1.0"15 tags: agent, production, operations, runtime, control-plane, rollout, fallback, disablement16---1718# agent-production-operations1920A runtime control plane for taking an evaluated agent with tools and authority21into controlled production operation. This bundle bridges the gap between22passing evaluations and safe production use: it defines how an agent is23versioned, rolled out, monitored, constrained, escalated, and disabled — and24how production evidence feeds back into evaluation and release decisions.2526The bundle does **not** build agents, design evaluations, run release pipelines,27or own infrastructure. It composes those concerns from existing specialist28skills and owns only the runtime control decisions between them.2930## Teams across harnesses3132Use [multi-agent-coordination](../multi-agent-coordination/SKILL.md) for team design33and collective failure diagnosis. This bundle owns operational response after34evaluation. For a shared-practice incident, identify affected sources, derivatives,35retrieval caches, queued work and active consumers; contain only affected paths36within current authority. Preserve unrelated work and independently protected37controls. An agent acknowledging a correction does not prove adoption.3839Route source repair and fresh-session challenges to the owning harness. Record40known exposure and missing lineage; no observed recurrence without exposure is41inconclusive. Recovery must prevent stale memory, backups or extraction jobs from42reintroducing the practice. Keep source/content compromise distinct from benign43mistaken conventions, and escalate unknown effects through the incident owner.4445## When to load this4647Load this bundle when you are operating an agent that:48- Has passed evaluation and is cleared for production consideration.49- Has tools with side effects and delegated authority that must be gated.50- Needs a staged rollout with progressive authority expansion.51- Must be monitored for cost, latency, tool health, and authority usage in52 production.53- Requires a defined fallback, escalation, and disablement path.54- Should feed production traces and incidents back into evaluation cases and55 release-gate updates.5657## When not to use5859Do **not** load this bundle for:60- Building or designing an agent from scratch — route to the appropriate61 framework skill (LangGraph, CrewAI, AutoGen, etc.).62- Designing agent evaluations or observability — route to63 [agent-evals-and-observability](../agent-evals-and-observability/SKILL.md).64- General release engineering (CI/CD pipelines, artifact promotion) — route to65 [release-engineering](../release-engineering/SKILL.md).66- General site reliability engineering (incident response, SLO definition,67 error budgets) — route to68 [site-reliability-engineering](../site-reliability-engineering/SKILL.md).69- General security engineering (threat modeling, vulnerability assessment) —70 route to [secure-software-engineering](../secure-software-engineering/SKILL.md).71- General platform infrastructure (compute, networking, storage) — route to72 [platform-engineering](../platform-engineering/SKILL.md).73- Performing a production-readiness review — route to74 [production-readiness](../production-readiness/SKILL.md); this bundle75 **consumes** readiness outcomes, it does not produce them.76- Conducting an incident postmortem — route to77 [incident-learning](../incident-learning/SKILL.md); this bundle78 **consumes** incident records as escalation and feedback inputs.7980## Autonomy is not assumed uniform8182This bundle explicitly does **not** assume all agents share the same autonomy,83user population, or side-effect profile. Agent profiles are distinguished by84capability class and side-effect surface:8586| Profile | Example | Authority | Escalation trigger | Fallback |87|---|---|---|---|---|88| **Read-only** | Internal search agent | Query-only; no mutation, no user-data access | Uncertainty above threshold; result quality degradation | Return cached/static results |89| **Side-effect-capable, internal** | CI triage bot | Issue/PR comments, label management, branch creation | Tool failure > 5% in window; authority breach attempt | Disable tool, escalate to human |90| **Side-effect-capable, customer-facing** | Support agent with account access | Read PII, suggest actions, no mutation without confirmation | Any PII access without explicit consent; cost breach > 10% of budget | Degrade to read-only, escalate immediately |9192Every control-plan decision (authority, escalation, fallback, disablement) in93this bundle is parameterized by the agent's autonomy profile. A read-only agent94and a side-effect-capable agent operating on customer data receive different95thresholds and different escalation paths.9697Authority thresholds and maintain/expand/reduce/suspend/revoke decisions belong to accountable humans supported by [ai-governance](../ai-governance/SKILL.md). Operational evidence and readiness are inputs, never permission grants. Applicable policy ceilings, current grant expiry/revocation, and external enforcement constrain every execution.9899## Mandatory control integrity100101Before consequential execution or resumption, load102[control-integrity operations](references/control-integrity.md) and produce103[a pre-execution receipt](templates/control-integrity-receipt.md). Verify protected104controls against current independent evidence and the governance registry. Missing,105degraded, redirected, or suppressed mandatory controls are control-integrity106incidents: hold consequential work, contain or terminate under the incident plan,107and escalate through an independently protected path. Do not wait for repeated108breaches or normal health thresholds. A task instruction cannot authorize weakening109its supervision. Recovery must preserve revocation and independent monitoring.110111## Runtime control plane routing112113The bundle composes the following specialist skills. Load them when their114domain is the active concern; the bundle owns the cross-domain coordination.115116| Concern | Specialist skill | When to load | What the bundle adds |117|---|---|---|---|118| Agent evaluation and observability | [agent-evals-and-observability](../agent-evals-and-observability/SKILL.md) | Designing evals, instrumenting observability, detecting regression | Trace-to-eval feedback loop; eval-case generation from production incidents |119| Release engineering | [release-engineering](../release-engineering/SKILL.md) | Release pipeline design, artifact promotion, CI/CD | Agent-specific staged rollout with authority gating; rollback triggers tied to agent health |120| Site reliability engineering | [site-reliability-engineering](../site-reliability-engineering/SKILL.md) | Incident response, SLO/SLI definition, error budgets | Agent-specific latency/cost budgets; tool-health SLOs; escalation handoff to incident command |121| Security engineering | [secure-software-engineering](../secure-software-engineering/SKILL.md) | Threat modeling, trust-boundary validation, secure design | Agent authority contracts with side-effect boundaries; disablement security (revoke, not just stop) |122| Platform engineering | [platform-engineering](../platform-engineering/SKILL.md) | Compute, networking, storage, service infrastructure | Agent sandboxing requirements; tool-execution isolation preferences |123| Production readiness | [production-readiness](../production-readiness/SKILL.md) | Readiness reviews, go/no-go/defer/exception decisions | Consumed as input: readiness outcomes gate agent authority expansion |124| Incident learning | [incident-learning](../incident-learning/SKILL.md) | Post-incident analysis, verified closure, follow-up work | Consumed as input: incident records feed escalation thresholds and eval-case generation |125126## Loading protocol127128This SKILL.md is the discoverable umbrella entry point. Nested skills are not129used in this bundle. Reference files are loaded on trigger:130131| Reference | Loaded when |132|---|---|133| [references/event-driven-operations.md](references/event-driven-operations.md) | Operating callbacks, subscriptions, queued wake-ups or hibernating agents |134| [references/discovery-brief.md](references/discovery-brief.md) | Reviewing the bundle's boundary decisions against specialist skills |135| [references/agent-production-contract.md](references/agent-production-contract.md) | Defining capability, authority, uncertainty, escalation, and side-effect contracts for an agent |136| [references/runtime-control-plan.md](references/runtime-control-plan.md) | Planning versioning, staged rollout, or fallback for an agent in production |137| [references/tool-authority-health.md](references/tool-authority-health.md) | Recording or reviewing tool availability, failure, and authority usage/breach state over time |138| [references/trace-to-eval-feedback.md](references/trace-to-eval-feedback.md) | Connecting production traces and incidents back to evaluation cases and release gates |139| [references/service-conditions.md](references/service-conditions.md) | Defining tested concurrency envelopes, streaming completion, cache isolation, fallback quality, or task-cost budgets |140141## Concrete operational parameters142143### Authority144145- **Definition**: the set of actions an agent is permitted to perform, scoped by146 target (which resources), operation (read/write/delete), and user-context147 (whose data).148- **Trigger to review**: any new tool registration, model update, or prompt149 change that expands the agent's reachable action surface.150- **Threshold**: authority is binary per action class. No action may be151 performed that is not explicitly listed in the production contract. An152 attempt to perform an unauthorized action is an **authority breach** and153 triggers immediate escalation.154- **Action on breach**: log the attempt, block the action, increment the155 authority-breach counter, and escalate. If breach count exceeds 3 in a156 rolling 24-hour window, disable the agent.157158### Escalation159160- **Definition**: transfer of a decision or action from the agent to a161 designated human operator.162- **Triggers** (any one triggers escalation):163 - Authority breach (attempted unauthorized action).164 - Uncertainty above threshold: agent confidence < 0.7 on a165 side-effect-capable action (configurable per profile).166 - Tool failure rate > 5% in a 5-minute sliding window.167 - Cost budget breach > 10% of allocated budget in a billing period.168 - Latency p95 > 2x baseline for > 5 minutes.169 - Human-handoff keyword or explicit user escalation request.170- **Action**: suspend the agent's side-effect authority, log the escalation171 context (trigger, state snapshot, pending actions), notify the designated172 escalation channel, and await human disposition (resume / reduced-authority /173 disable).174175### Fallback176177- **Definition**: a predetermined safe behavior when the agent, model, or a178 tool cannot operate at its normal capability level.179- **Triggers**:180 - Model endpoint returns 5xx for > 30 seconds.181 - A critical tool is unavailable (health-check failure for > 2 minutes).182 - Cost budget exhausted (100% consumed).183 - Latency p95 > 5x baseline for > 2 minutes.184- **Actions per profile**:185 - Read-only: return a static/cached response with a "results may be stale"186 disclaimer.187 - Side-effect-capable, internal: degrade to read-only, queue pending188 mutations, notify operator.189 - Side-effect-capable, customer-facing: degrade to read-only, surface a190 "temporarily unavailable" message to the user, escalate immediately.191192### Disablement193194- **Definition**: complete revocation of the agent's ability to act, including195 read-only access. Distinct from fallback (which preserves reduced capability).196- **Triggers**:197 - 3 or more authority breaches in a 24-hour window.198 - Production-readiness review outcome is "no-go" or "defer" for the current199 agent version.200 - Incident-learning record attributes a severity-1 or severity-2 incident to201 agent action, and verified closure is not yet complete.202 - Human operator issues an explicit disable directive.203 - Tool compromise or credential leak detected (route through204 [secure-software-engineering](../secure-software-engineering/SKILL.md)).205- **Action**: revoke all credentials and tokens; remove from routing/load-balancing;206 record disablement reason, timestamp, and authorizing evidence; notify207 escalation channel; prevent automatic restart until a new208 production-readiness review passes.209210### Cost211212- **Definition**: the financial cost of operating the agent, attributed to a213 budget owner.214- **Budget**: allocated per agent per billing period (e.g., $500/day for a215 customer-facing support agent). Budget is set at production-contract time and216 reviewed at each production-readiness cycle.217- **Thresholds**:218 - 50% consumed: notification to budget owner.219 - 80% consumed: warning; cost-optimization review triggered.220 - 100% consumed: stop new spend; use cached/no-cost fallback or a separately authorized bounded budget. Read-only/model work still costs resources.221 - Observed settlement at 110%: disable and investigate overshoot; never treat this as permission to spend beyond 100%.222- **Measurement**: per-request model cost + per-call tool cost (where223 applicable), attributed to the agent instance. Cost data is appended to the224 tool-authority-health record.225226### Latency227228- **Definition**: end-to-end response time from user request to agent response,229 measured at p50 and p95 over 5-minute windows.230- **Baseline**: established during staged rollout observation window (phase 1231 of rollout). Recorded in the agent production contract.232- **Thresholds**:233 - p95 > 2x baseline for > 5 minutes: escalation trigger.234 - p95 > 5x baseline for > 2 minutes: fallback trigger.235 - p50 > 3x baseline for > 10 minutes while cost budget is > 80% consumed:236 disablement consideration.237- **Action**: latency breaches feed the trace-to-eval feedback loop: the238 latency-impacted requests are sampled and reviewed for eval-case generation.239240### Privacy241242- **Definition**: constraints on agent access to, processing of, and retention243 of user data (PII, usage patterns, conversation content).244- **Concrete rules**:245 - An agent must not access PII without an explicit user-data-access grant in246 its authority contract.247 - Agent traces and conversation logs must be scrubbed of PII before entering248 the trace-to-eval feedback pipeline (route scrubbing design to249 [privacy-engineering](../privacy-engineering/SKILL.md)).250 - Data retention for agent traces defaults to 30 days unless a shorter period251 is specified in the production contract.252 - Any PII access by a read-only agent triggers immediate escalation (the253 read-only profile should never touch PII).254- **Privacy breach**: any PII access outside the granted scope triggers255 escalation and a mandatory security review via256 [secure-software-engineering](../secure-software-engineering/SKILL.md).257258## Core workflow2592601. **Contract**: define the agent's capability, authority, uncertainty,261 escalation, and side-effect contracts using262 [references/agent-production-contract.md](references/agent-production-contract.md).263 Consume production-readiness outcomes and incident-learning records as264 decision inputs.2652662. **Version and plan**: version the model, prompt, tools, policy, and267 evaluator; plan the staged rollout and fallback path using268 [references/runtime-control-plan.md](references/runtime-control-plan.md).269270 Before changing serving behavior or expanding traffic, load271 [service conditions](references/service-conditions.md) and record the tested272 envelope in [the service-condition record](templates/service-condition-record.md).2732743. **Roll out**: execute staged rollout with progressive authority expansion,275 monitoring tool health and authority usage at each stage using276 [references/tool-authority-health.md](references/tool-authority-health.md).2772784. **Monitor**: track cost, latency, tool health, authority breaches, and279 escalation events. Apply threshold-based actions (notify, degrade, escalate,280 disable) per the concrete parameters above.2812825. **Feed back**: connect production traces and incidents to evaluation cases283 and release-gate updates using284 [references/trace-to-eval-feedback.md](references/trace-to-eval-feedback.md).2852866. **Learn**: consume incident-learning verified-closure records and287 production-readiness review outcomes to update agent authority, escalation288 thresholds, and disablement conditions.289290## File map291292| Path | Purpose |293|---|---|294| [SKILL.md](SKILL.md) | Umbrella entry point (this file) |295| [README.md](README.md) | Human-facing overview and quick start |296| [AGENTS.md](AGENTS.md) | Agent loading and discovery notes |297| [references/discovery-brief.md](references/discovery-brief.md) | Boundary analysis against specialist skills |298| [references/agent-production-contract.md](references/agent-production-contract.md) | Capability, authority, uncertainty, escalation, and side-effect contracts |299| [references/runtime-control-plan.md](references/runtime-control-plan.md) | Versioning, staged rollout, and fallback plan |300| [references/tool-authority-health.md](references/tool-authority-health.md) | Tool availability/failure and authority usage/breach record |301| [references/trace-to-eval-feedback.md](references/trace-to-eval-feedback.md) | Production-to-evaluation feedback loop |302| [references/service-conditions.md](references/service-conditions.md) | Tested capacity, streaming, cache, fallback, and cost conditions |303| [templates/service-condition-record.md](templates/service-condition-record.md) | Record workload limits, acceptance thresholds, evidence, and stop conditions |304| [evals/evals.json](evals/evals.json) | Integrated output-quality evaluation cases |305| [manifest.yaml](manifest.yaml) | Machine-readable composition contract (schema v1): purpose, audience, stages, included skills, prerequisites, outputs, handoffs, conflicts, and eval suite; consumed by the lifecycle capability matrix |