Inspiration
Long-running AI codebases rarely fail only because code is wrong. They also fail when agents lose track of what already exists: roadmaps outlive implementations, design documents are mistaken for shipped features, and parallel agents rebuild equivalent capabilities.
SAEE is a pre-existing Digital Biosphere Evolution Engine. During OpenAI Build Week, we extended it with a bounded developer tool—the SAEE Evolution Capability Router—to make capability evolution discoverable, composable, and safer for coding agents.
What it does
- Gives Codex and other coding agents one canonical, machine-readable capability inventory.
- Classifies every capability as
implemented,partial,design_only,missing,deprecated, orsuperseded. - Resolves public aliases to one canonical CLI, MCP, HTTP, or Python entrypoint.
- Blocks duplicate builds when an equivalent capability already exists.
- Separates local code, synthetic validation, external integration, customer validation, and production readiness.
- Fails closed when routing is ambiguous or when a requested interface does not exist.
What was built during Build Week
SAEE existed before the submission period. The judged extension starts after the official submission window opened. From commit 9f74d153 through f6ac41f4, Codex helped add a canonical capability inventory, runtime loader, CLI routing commands, schema validation, duplicate-build gates, governance registries, and dogfooding evidence. That range changes 74 files with 6,850 insertions and 40 deletions. Pre-existing SAEE simulation and product surfaces are context, not claimed as Build Week work.
How we built it
GPT-5.6 in Codex was used to inventory conflicting truth surfaces, trace aliases to implementations, propose fail-closed contracts, generate negative cases, and review staged-truth language. Human decisions fixed the constitutional boundary: the router strengthens SAEE's Evolutionary Archive / Rollback Immune System but does not replace the Digital Biosphere Evolution Engine core.
The implementation is deterministic Python with JSON contracts. It can be tested offline and does not execute unknown repositories, contact external systems, or expand permissions.
Challenges we ran into
The hardest problem was not adding another tool; it was proving that another tool was not already present. Historical roadmap fields, public metadata, runtime adapters, and documentation had drifted. We solved this by making capability-package/manifest.json#canonical_inventory the only capability fact source and validating every projection against it.
A second challenge was preventing truthful local results from becoming inflated claims. The tool therefore reports missing capabilities explicitly and keeps production_ready=false.
Accomplishments that we're proud of
A clean snapshot at f6ac41f4 passes:
- canonical inventory validation: 9/9 capabilities, 16/16 negative cases, and 5/5 deterministic runs;
- capability progress ledger validation: 9/9 status projections, with duplicate-build prevention enabled;
- governance registry validation: 6/6 registries and 4/4 schemas.
The inventory intentionally reports three implemented capabilities, one partial capability, one design-only capability, and four missing capabilities.
What we learned
Agent-readable contracts are a product surface, not a documentation afterthought. A useful coding agent needs explicit canonical entrypoints, non-claims, lifecycle states, and deterministic validators more than another prose roadmap.
What's next for SAEE Evolution Capability Router
After the hackathon, the router can support the constitutionally controlled integration of SAEE and Agent Evidence. Public MCP transport, external interoperability, customer validation, and production readiness remain future gates—not current claims.
Truth boundary
build_week_commit_range=9f74d153..f6ac41f4
clean_snapshot=f6ac41f4b068377e7778e8c3d83b99bd8382debc
public_judging_branch_commit=486a50bd272e93865e502fc1aa7d8fb0ee024eda
local_validation=pass
synthetic_validation=pass
public_mcp_endpoint_available=false
external_interoperability_validated=false
customer_validation=false
production_ready=false
Log in or sign up for Devpost to join the conversation.