Harness engineering
A reference corpus, not a dependency. It holds a thesis and playbooks for improving
agent output by shaping the environment around a fixed model and agent — not code a
project imports or vendors. Ryan Lopopolo's work, bundled here under CC BY 4.0; see
COPYING.md and LICENSE.
When to reach for it
- Only on a genuinely unresolved decision. Read the target repo's own
instructions, domain model, and precedent first. Consult this corpus for the
specific concern local evidence leaves open. Don't preload it for routine work.
- Target-local truth always wins. The corpus can sharpen a decision; it never
overrides the target repo's contracts, authority, or conventions. Adapt the
applicable idea — don't copy its file layout, policies, or fixtures.
- Treat it as read-only. It is reference material, not a working tree.
Working loop
- Read the target's available instructions and architecture first. For a repository,
begin with its agent guide.
- Inspect the target's actual domain model, manifests, tools, tests, permissions,
operating history, and neighboring implementations. Find the local owner or
precedent for the concern before retrieving general guidance.
- Name the governing decision that local evidence leaves unresolved. Stop here when
none remains. Otherwise pick one primary route below and read that thesis
README.md. Add a second route only for a genuinely distinct concern.
Context routing
Route an unresolved decision to one thesis section under
reference/docs/:
| Unresolved concern |
Route |
| What the agent can see and when |
just-in-time-context |
| Whether a tool's surface is readable to the agent |
tool-legibility |
| What the agent is allowed to decide or do |
authority |
| Modeling the domain so requirements are recoverable |
domain-modeling |
| Showing the outcome actually holds |
proof |
| Closing the loop from a run back into the harness |
feedback |
| Systems that survive their maintainers |
durable-systems |
| Keeping a repo healthy over time |
continuous-maintenance |
| Whether the harness is actually working |
effectiveness |
| Scoping the agent to a whole job, not a fragment |
whole-job |
| Holding the worker fixed while changing its environment |
fixed-worker |
| Getting work the last mile into production |
last-mile-deployment |
| Tracing a claim back to its source |
lineage |
Application routing
For applying the practice rather than deciding a point of design, read
reference/playbooks/README.md before selecting a
procedure:
- Improve the harness around one observed job —
improve-harness.md, read after the
target's instructions and before collecting a baseline or making a change. Use the
thesis routing above inside its gap-classification step.
- Assess a repository broadly —
repository-review.md, read before
beginning the review.
- Make a comparative, causal, or longitudinal claim —
reference/evals/README.md, read before running
conditions.
To learn or teach the practice as a whole, start at
reference/README.md and the thesis index in
reference/docs/README.md.
Provenance
reference/sources/ holds the raw source material the
prose draws on. Consult it to check a claim against its origin; it is provenance, not
guidance. The CC BY 4.0 grant excludes third-party quotations, images, screenshots,
embeds, logos, and trademarks identified in the sources — those remain with their
rightsholders.
1---2name: engineering3description: Harness engineering4---56# Harness engineering78A reference corpus, not a dependency. It holds a thesis and playbooks for improving9agent output by shaping the environment around a fixed model and agent — not code a10project imports or vendors. Ryan Lopopolo's work, bundled here under CC BY 4.0; see11[COPYING.md](COPYING.md) and [LICENSE](LICENSE).1213## When to reach for it14151. **Only on a genuinely unresolved decision.** Read the target repo's own16 instructions, domain model, and precedent first. Consult this corpus for the17 specific concern local evidence leaves open. Don't preload it for routine work.182. **Target-local truth always wins.** The corpus can sharpen a decision; it never19 overrides the target repo's contracts, authority, or conventions. Adapt the20 applicable idea — don't copy its file layout, policies, or fixtures.213. **Treat it as read-only.** It is reference material, not a working tree.2223## Working loop24251. Read the target's available instructions and architecture first. For a repository,26 begin with its agent guide.272. Inspect the target's actual domain model, manifests, tools, tests, permissions,28 operating history, and neighboring implementations. Find the local owner or29 precedent for the concern before retrieving general guidance.303. Name the governing decision that local evidence leaves unresolved. Stop here when31 none remains. Otherwise pick **one** primary route below and read that thesis32 `README.md`. Add a second route only for a genuinely distinct concern.3334## Context routing3536Route an unresolved decision to one thesis section under37[reference/docs/](reference/docs/README.md):3839| Unresolved concern | Route |40| --- | --- |41| What the agent can see and when | [just-in-time-context](reference/docs/just-in-time-context/README.md) |42| Whether a tool's surface is readable to the agent | [tool-legibility](reference/docs/tool-legibility/README.md) |43| What the agent is allowed to decide or do | [authority](reference/docs/authority/README.md) |44| Modeling the domain so requirements are recoverable | [domain-modeling](reference/docs/domain-modeling/README.md) |45| Showing the outcome actually holds | [proof](reference/docs/proof/README.md) |46| Closing the loop from a run back into the harness | [feedback](reference/docs/feedback/README.md) |47| Systems that survive their maintainers | [durable-systems](reference/docs/durable-systems/README.md) |48| Keeping a repo healthy over time | [continuous-maintenance](reference/docs/continuous-maintenance/README.md) |49| Whether the harness is actually working | [effectiveness](reference/docs/effectiveness/README.md) |50| Scoping the agent to a whole job, not a fragment | [whole-job](reference/docs/whole-job/README.md) |51| Holding the worker fixed while changing its environment | [fixed-worker](reference/docs/fixed-worker/README.md) |52| Getting work the last mile into production | [last-mile-deployment](reference/docs/last-mile-deployment/README.md) |53| Tracing a claim back to its source | [lineage](reference/docs/lineage/README.md) |5455## Application routing5657For applying the practice rather than deciding a point of design, read58[reference/playbooks/README.md](reference/playbooks/README.md) before selecting a59procedure:6061- **Improve the harness around one observed job** —62 [improve-harness.md](reference/playbooks/improve-harness.md), read after the63 target's instructions and before collecting a baseline or making a change. Use the64 thesis routing above inside its gap-classification step.65- **Assess a repository broadly** —66 [repository-review.md](reference/playbooks/repository-review.md), read before67 beginning the review.68- **Make a comparative, causal, or longitudinal claim** —69 [reference/evals/README.md](reference/evals/README.md), read before running70 conditions.7172To learn or teach the practice as a whole, start at73[reference/README.md](reference/README.md) and the thesis index in74[reference/docs/README.md](reference/docs/README.md).7576## Provenance7778[reference/sources/](reference/sources/README.md) holds the raw source material the79prose draws on. Consult it to check a claim against its origin; it is provenance, not80guidance. The CC BY 4.0 grant excludes third-party quotations, images, screenshots,81embeds, logos, and trademarks identified in the sources — those remain with their82rightsholders.