EGO: Elephant, Goldfish, Owl
EGO adds a third role to Google's Elephant-Goldfish Model (EGM). In EGM the Elephant drafts a design document and Goldfish reviewers, who remember nothing, attack it. The Elephant then decides which attacks stand, so it grades its own work.
EGO adds the Owl, a judge that remembers only the argument. It rules on disputes blind: it never learns which side wrote the document. Questions of value go to the person.
What it changes
| EGM | EGO |
|---|---|
| The Elephant labels each finding CONTEST or PROPOSE FIX and that label stands. | The Elephant answers FIX, CONTEST or REFRAME. The Goldfish replies once. The Owl rules on each dispute. |
| Findings are deduplicated and sorted by severity. | ego-tensions finds pairs of findings that ask for opposite changes, and sends them to the person. |
| Two outcomes. | Four verdicts: SUSTAIN, OVERRULE, REFRAME, ESCALATE. |
| At the re-audit cap, leftovers become "accepted trade-offs". | Every leftover Blocker or Major ends as a fix, a recorded ruling, or a person's answer. |
| Settled arguments return each round. | Rulings go into the document's ## Decisions section, which later Goldfish read. |
The Owl follows five rules: rule only after both sides speak, positions are blind and in random order, a position wins only on quoted evidence, value calls escalate, and the Owl never writes the fix.
Install
EGO runs in the same harnesses as EGM: Claude Code, Claude Cowork, Codex, Antigravity and ChatGPT. Install EGM first in the same harness. Claude Code:
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm
claude plugin marketplace add radiator-engineering/ego
claude plugin install ego@ego
Codex:
codex plugin marketplace add google/elephant-goldfish-model && codex plugin add egm@egm
codex plugin marketplace add radiator-engineering/ego && codex plugin add ego@ego
The others are in docs/harnesses.md.
Use
/ego:ego-review <doc>(Claude; other harnesses pick theego-reviewskill): EGM's staged review with the EGO changes. It reads EGM's own review skill from the installed EGM, so EGM updates carry over./ego:ego-resolve: resolve any Goldfish report (for example one written by/egm:egm-review) with the Owl.- Reviewing a document someone else owns: pass
Mode: suggestandOwner: <name>to either skill. EGO leaves the document untouched. It writessuggested_edits.mdandproposed_decisions.mdfor the owner and addresses the value calls to them.
Drafting and brainstorming stay in EGM: /egm:egm-brainstorm,
/egm:egm-draft.
Layout
| Path | What it is |
|---|---|
agents/owl.md | The Owl (ego-owl). |
agents/tensions.md | The tension finder (ego-tensions). |
.claude-plugin/agents/ | Claude adapters: read-only tools, no CLAUDE.md. |
skills/ego-resolve/ | Resolution with the Owl; replaces egm-resolve. |
skills/ego-review/ | EGM's review plus four overrides. |
scripts/egm-root.sh | Finds the installed EGM. |
scripts/blind-check.py | Strips authorship leaks from Owl positions. |
docs/harnesses.md | Install, roots and agent names per harness. |
docs/measures.md | How to tell whether the Owl helps. |
Limits
ego-reviewfollows EGM'segm_review/SKILL.mdby reference. If EGM renames that file or theegm-resolvehand-off, the overrides need an update.- Not yet built: an antithesis lens for brainstorming, and an Owl ruling when two instances of the same reviewer disagree.
- Not yet measured. See
docs/measures.md.
License
Apache-2.0. EGO contains no EGM source; it calls the installed EGM plugin.