diff --git a/docs/.vitepress/config.ts b/docs/.vitepress/config.ts
index 0b0f3c581b..573f98c511 100644
--- a/docs/.vitepress/config.ts
+++ b/docs/.vitepress/config.ts
@@ -213,6 +213,7 @@ export default defineConfig({
             { text: "Retro", link: "/agents/retro" },
             { text: "Prioritize", link: "/agents/prioritize" },
             { text: "Default vs. Custom", link: "/agents/topics/default-vs-custom" },
+            { text: "Escalation Ladder", link: "/agents/topics/escalation-ladder" },
           ],
         },
         {
diff --git a/docs/agents/topics/default-vs-custom.md b/docs/agents/topics/default-vs-custom.md
index 984b6734da..5bffc8fd32 100644
--- a/docs/agents/topics/default-vs-custom.md
+++ b/docs/agents/topics/default-vs-custom.md
@@ -87,6 +87,8 @@ for a stated purpose, that override does not make the agent derived.
 
 ## See also
 
+- [Escalation ladder](escalation-ladder.md) — prove-it path before
+  deriving or replacing a core agent
 - [Agents reference](../) — default agent documentation and extension points
 - [Configuring agent behavior](../../guides/user/customizing-agents.md) — harness
   configuration and layered content resolution
diff --git a/docs/agents/topics/escalation-ladder.md b/docs/agents/topics/escalation-ladder.md
new file mode 100644
index 0000000000..76a8ad0607
--- /dev/null
+++ b/docs/agents/topics/escalation-ladder.md
@@ -0,0 +1,194 @@
+# Escalation ladder for changing agent behavior
+
+> **Status: Proposed.** Level names and evidence-to-escalate criteria are a
+> working proposal pending maintainer agreement on
+> [#5372](https://github.com/fullsend-ai/fullsend/issues/5372). Treat this as
+> guidance under review, not settled project policy.
+
+Before replacing or deriving from a core agent, prove that lighter options
+cannot meet your needs. This page defines the escalation path — start at
+Level 1 and move up only when you can show why the current level is
+insufficient.
+
+This ladder applies to **core agent roles that fullsend already ships**
+(triage, code, review, fix, retro, prioritize). If you are building an agent
+for a job fullsend does not cover — a release-notes generator, a
+compliance checker, a deployment orchestrator — skip ahead to
+[Bring Your Own Agent](../../guides/user/bring-your-own-agent.md). There is
+no existing default to exhaust first.
+
+## The four levels
+
+```
+Level 1 — Configure           Least invasive. No code leaves your repo.
+Level 2 — Contribute          Fix the gap in the default agent for everyone.
+Level 3 — Derive              Inherit the default, replace identity components.
+Level 4 — Replace             Build a parallel agent for the same role.
+```
+
+Each level increases maintenance burden. A derived agent inherits upstream
+improvements via `base` but must track breaking changes in the parent
+harness. A replacement agent inherits nothing — every upstream improvement
+must be replicated independently.
+
+## Level 1: Configure the default agent
+
+Use the extension points the default agent was designed for. These keep you
+in [configured-default territory](default-vs-custom.md) and require no
+changes to the agent's identity-defining fields. Some rows need a thin
+`base`-composed harness (still Level 1); others need no harness at all.
+
+| Extension point | What it does | Guide |
+|---|---|---|
+| `AGENTS.md` | Project-wide instructions for all agents — code style, test commands, architecture rules, domain context. No harness required. | [Customizing with AGENTS.md](../../guides/user/customizing-with-agents-md.md) |
+| Repo skills (`.agents/skills/`) | Domain-specific knowledge for individual agents — linting rules, deployment checklists, label glossaries. No harness required. | [Customizing with Skills](../../guides/user/customizing-with-skills.md) |
+| Documented env vars | Per-agent tuning knobs (e.g., `REVIEW_FINDING_SEVERITY_THRESHOLD`). No harness required. | Each agent's [reference page](../) |
+| `env:` in harness | Add environment variables without changing the agent's identity. Needs a thin `base` harness. | [Harness field reference](../../guides/user/bring-your-own-agent.md#harness-field-reference) |
+| `skills:` in harness | Add skills via `base` composition — concatenated with the base agent's skill list. Needs a thin `base` harness. | [Configuring existing agents](../../guides/user/bring-your-own-agent.md#configuring-existing-agents) |
+| `plugins:` in harness | Add language-server plugins. Needs a thin `base` harness. | [Harness field reference](../../guides/user/bring-your-own-agent.md#harness-field-reference) |
+| `host_files:` in harness | Inject additional files into the sandbox. Needs a thin `base` harness. | [Harness field reference](../../guides/user/bring-your-own-agent.md#harness-field-reference) |
+| Sandbox image layers | Base your image on the default, add tools; point the harness `image:` field at it. Needs a thin `base` harness. | [Extending the sandbox image](../../guides/user/customizing-agents.md#extending-the-sandbox-image) |
+
+**Evidence to escalate:** show that no combination of these extension points
+produces the behavior you need. Concrete evidence includes:
+
+- A failing test case or eval run where the agent consistently gets the
+  wrong answer despite correct `AGENTS.md` instructions
+- A skill that cannot influence the agent's behavior because the gap is in
+  the system prompt, not in domain knowledge
+- An env var or config knob that does not exist for the behavior you need to
+  change
+
+## Level 2: Contribute to the default agent
+
+If Level 1 cannot close the gap, check whether the gap is general — would
+other users benefit from the same change? If yes, the fix belongs in the
+default agent, not in a private fork.
+
+Contributions include:
+
+- **New extension points** — a new env var, a new skill hook, a new
+  documented configuration knob
+- **Prompt improvements** — clarifications, better instructions, additional
+  decision criteria in the agent definition
+- **Script changes** — pre/post script logic that handles a broader set of
+  cases
+- **Skill improvements** — better procedures, additional steps, broader
+  coverage in a built-in skill
+
+**How to contribute:** file an issue or open a PR against the default agent.
+If the change is in the [fullsend-ai/fullsend](https://github.com/fullsend-ai/fullsend)
+repo (harness, scripts, skills, agent definitions), contribute there. If the
+agent definition or harness lives in
+[fullsend-ai/agents](https://github.com/fullsend-ai/agents), contribute to
+that repo instead. First-time contributors must be
+[vouched](../../../CONTRIBUTING.md) by a maintainer before opening a PR —
+unvouched PRs are auto-closed.
+
+**Evidence to escalate:** show that the improvement you need is specific to
+your team's workflow and would not benefit other users. Concrete evidence
+includes:
+
+- A proposed upstream change that was reviewed and rejected as too
+  org-specific
+- A behavior change that contradicts the default agent's documented charter
+  (e.g., you need the review agent to auto-merge, but the default review
+  agent is explicitly read-only)
+- A workflow that requires proprietary integrations (e.g., posting to an
+  internal Slack channel, querying a private API)
+
+## Level 3: Derive from the default agent
+
+Inherit the default harness via `base` and replace only the components that
+must differ. This is a [derived agent](default-vs-custom.md) — it tracks
+upstream improvements for everything you did not override, but the
+components you replaced are now your responsibility.
+
+Derived agents use `base` composition to inherit the default harness and
+override identity-defining fields:
+
+```yaml
+base: https://raw.githubusercontent.com/fullsend-ai/fullsend/<sha>/internal/scaffold/fullsend-repo/harness/code.yaml#sha256=abc...
+
+# Override identity-defining components:
+agent: agents/my-code-agent.md          # Custom system prompt
+post_script: scripts/post-my-code.sh    # Custom post-processing
+slug: my-org-code                       # Custom identity
+```
+
+See [Configuring existing agents](../../guides/user/bring-your-own-agent.md#configuring-existing-agents)
+for the full pattern and
+[Classification by harness field](default-vs-custom.md#classification-by-harness-field)
+for which fields cross the derived threshold.
+
+**What you maintain:**
+
+- Your custom system prompt, scripts, or validation loop
+- Compatibility with upstream `base` harness changes (field additions,
+  schema changes, script interface changes)
+- Your own testing and evaluation for the overridden components
+
+**Evidence to escalate:** show that `base` inheritance cannot support your
+use case. Concrete evidence includes:
+
+- A fundamental incompatibility with the base harness's script interface
+  (e.g., the base post-script expects an output format your agent does not
+  produce, and you cannot adapt either side)
+- A need to remove or replace a provider the base harness already declares —
+  `providers:` entries are concatenated (base + child), not subtractive.
+  Sandbox policy and image can already be fully overridden at Level 3 via a
+  scalar field replacement
+
+## Level 4: Replace the role entirely
+
+Build a custom agent from scratch that fills the same role as a default
+agent. This is the heaviest option — you inherit nothing from upstream and
+must maintain the full harness, agent definition, scripts, skills, and
+testing independently.
+
+See [Bring Your Own Agent](../../guides/user/bring-your-own-agent.md) for
+the end-to-end guide.
+
+**When this is appropriate:**
+
+- The default agent's architecture is fundamentally incompatible with your
+  requirements (not just its prompt or scripts)
+- You have passed through Levels 1–3 and documented why each is
+  insufficient
+- You have the capacity to maintain a parallel implementation long-term
+
+Register your replacement agent with the same role name in `config.yaml` —
+config-registered agents take precedence over built-in agents on name
+collision. See
+[Registering your agent](../../guides/user/bring-your-own-agent.md#registering-your-agent).
+
+## Prove-it checklist
+
+Before creating a derived or replacement agent for a core role, confirm:
+
+- [ ] **Level 1 exhausted.** You tried `AGENTS.md`, repo skills, env vars,
+  harness `skills:` / `plugins:` / `host_files:`, and documented config
+  knobs. None of them address the gap.
+- [ ] **Level 2 considered.** The improvement is too org-specific to
+  contribute upstream, or an upstream contribution was proposed and
+  rejected.
+- [ ] **Level 3 evaluated** (if jumping to Level 4). `base` inheritance
+  cannot support your use case, and you can articulate why.
+- [ ] **Evidence documented.** You have specific examples — failed eval
+  runs, rejected upstream PRs, architectural constraints — not just a
+  preference for a different approach.
+- [ ] **Maintenance plan.** You have a plan for keeping the derived or
+  custom agent up to date as the default evolves.
+
+## See also
+
+- [Default, derived, and custom agents](default-vs-custom.md) — how to
+  classify your agent after making changes
+- [Bring Your Own Agent](../../guides/user/bring-your-own-agent.md) —
+  building and registering custom agents
+- [Customizing with AGENTS.md](../../guides/user/customizing-with-agents-md.md)
+  — Level 1: project-wide instructions
+- [Customizing with Skills](../../guides/user/customizing-with-skills.md)
+  — Level 1: agent-specific domain knowledge
+- [Customizing agents](../../guides/user/customizing-agents.md) — harness
+  configuration and layered content resolution
diff --git a/docs/guides/user/bring-your-own-agent.md b/docs/guides/user/bring-your-own-agent.md
index 51c7904a5b..eb63210e6b 100644
--- a/docs/guides/user/bring-your-own-agent.md
+++ b/docs/guides/user/bring-your-own-agent.md
@@ -585,4 +585,5 @@ When configured with `FALLBACK_MINT_URL`, the standalone mint serves custom role
 - [Configuring with AGENTS.md](customizing-with-agents-md.md) — repo-level instructions for all agents
 - [Configuring agent behavior](customizing-agents.md) — harness configurations and layered content resolution
 - [Default, derived, and custom agents](../../agents/topics/default-vs-custom.md) — when configuration crosses into custom agent territory
+- [Escalation ladder](../../agents/topics/escalation-ladder.md) — prove-it path before deriving or replacing a core agent
 - [Standalone mint](../infrastructure/standalone-mint.md) — custom agent roles and identity
diff --git a/docs/guides/user/building-custom-agents.md b/docs/guides/user/building-custom-agents.md
index c07653ec57..39e5b21289 100644
--- a/docs/guides/user/building-custom-agents.md
+++ b/docs/guides/user/building-custom-agents.md
@@ -9,8 +9,9 @@ This guide walks through creating a custom from-scratch agent on a per-repo
 fullsend installation.
 
 Before building from scratch, consider whether extending a default agent would
-meet your needs. You can use `base` inheritance to start from a default agent's
-harness and override only what differs — see
+meet your needs. Start with the
+[escalation ladder](../../agents/topics/escalation-ladder.md) to exhaust
+lighter options first, and see
 [Default, derived, and custom agents](../../agents/topics/default-vs-custom.md)
 for the distinction and when each approach makes sense.
 
diff --git a/docs/guides/user/customizing-agents.md b/docs/guides/user/customizing-agents.md
index ca3666fc52..5d02c3d3dc 100644
--- a/docs/guides/user/customizing-agents.md
+++ b/docs/guides/user/customizing-agents.md
@@ -360,6 +360,23 @@ Each agent role has its own identity, permissions, and purpose:
 
 ## Configuration Examples
 
+### Extending the sandbox image
+
+When `host_files` injection is not enough and you need additional packages or
+toolchains in the sandbox, build an image that extends the published base and
+point your harness `image:` field at it:
+
+```dockerfile
+FROM ghcr.io/fullsend-ai/fullsend-sandbox:latest
+RUN apt-get update && apt-get install -y --no-install-recommends rustc \
+  && rm -rf /var/lib/apt/lists/*
+```
+
+Use `ghcr.io/fullsend-ai/fullsend-code:latest` as the parent instead when you
+also need the Go toolchain. Then set `image:` in a thin `base`-composed
+harness (see [Configuring existing agents](bring-your-own-agent.md#configuring-existing-agents)).
+Pin the parent tag to a digest before CI use.
+
 ### Adding Executables
 
 The sandbox already has `/sandbox/workspace/bin` on its `PATH`. To make a
@@ -546,6 +563,7 @@ disables nothing because no agent has that harness name.
 
 - [Bring Your Own Agent](bring-your-own-agent.md) - Building and registering custom agents from scratch
 - [Default, derived, and custom agents](../../agents/topics/default-vs-custom.md) - When does configuration cross into derived or custom agent territory?
+- [Escalation ladder](../../agents/topics/escalation-ladder.md) - Prove-it path before deriving or replacing a core agent
 - [Getting Started](../getting-started/) - Initial setup
 - [Bugfix Workflow](bugfix-workflow.md) - How agents work together
 - [Standalone Mint](../infrastructure/standalone-mint.md) - Running your own mint with custom agent roles