Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding.
Ornith-1.0 spans the full parameter sizes including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. It achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks including:
✅Terminal-Bench 2.1(77.5)
✅SWE-Bench(82.4 on verified, 62.2 on pro, 78.9 on Multilingual)
✅NL2Repo(48.2)
✅SWE Atlas(41.2 on QnA, 42.6 RF, 39.1 TW)
✅ClawEval(77.1)
Post-trained on top of gemma4 and qwen3.5, Ornith-1.0 employs a novel self-improving training strategy in which reinforcement learning is used to generate not only solution rollouts, but also the task-specific scaffolds that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model generate higher-quality solutions in agentic coding.😎
All models are released under the MIT license, enabling full commercial and research use.
📖Tech Blog: https://t.co/qT9N2HYWFn
🤗Huggingface: https://t.co/PRrwqjeBtM
OpenAI engineer:
“90% of our engineers use Codex subagent harnesses + loops to ship code 5× faster.
close the agent loop. Give an agent a way to verify its own output.”
in 30-minute talk, an OpenAI engineer explains how to build self-improving agentic systems with harnesses.
Worth more than a $500 agentic course online.
Watch it today, then read how hedge funds apply loop engineering in trading.
This is mine. /goal
Operate as a continuous software quality and validation agent.
PHASE 1: FEATURE DISCOVERY
1. Analyse the entire codebase.
2. Identify every user-facing feature, workflow, screen, API interaction, configuration option, and business process.
3. For each feature:
- Create a unique Feature ID.
- Create a detailed user story.
- Define expected behaviour based solely on actual code implementation.
- Document edge cases.
- Document validation rules.
- Document dependencies.
- Document known assumptions.
4. Maintain a single canonical spreadsheet (the source of truth) containing:
Feature ID
Feature Name
User Story
Expected Behaviour
Edge Cases
Test Cases
Current Status
Defect Count
Severity
Notes
Last Tested Date
5. Continue discovery until no undocumented feature remains.
Exit Criteria:
- Every identifiable feature in the codebase exists in the spreadsheet.
- No screens, routes, workflows, or APIs remain undocumented.
--------------------------------------------------
PHASE 2: TEST GENERATION
For every feature:
1. Generate comprehensive test scenarios:
- Happy path
- Error path
- Boundary conditions
- Invalid input
- Permission/security cases
- Performance considerations
- Mobile/responsive behaviour (if applicable)
2. Add all test cases to the spreadsheet.
Exit Criteria:
- Every feature has at least one complete test suite.
- All major user journeys are covered end-to-end.
--------------------------------------------------
PHASE 3: EXECUTION
Execute every test case.
For every failure:
1. Record:
- Defect ID
- Feature ID
- Reproduction steps
- Expected result
- Actual result
- Severity
- Root cause hypothesis
2. Update spreadsheet immediately.
Exit Criteria:
- Every test case executed.
- Every defect documented.
--------------------------------------------------
PHASE 4: REMEDIATION
For each defect:
1. Investigate root cause.
2. Implement the smallest safe fix.
3. Verify fix locally.
4. Update defect status.
Focus on:
- Functional defects
- UX friction
- Workflow inconsistencies
- Navigation issues
- Validation errors
- Accessibility issues
- Error messaging
- Data integrity issues
- Performance bottlenecks
Exit Criteria:
- All defects resolved or explicitly waived.
--------------------------------------------------
PHASE 5: REGRESSION TESTING
1. Re-run every test case.
2. Re-run all major end-to-end user journeys.
3. Verify no regressions were introduced.
4. Update spreadsheet.
Exit Criteria:
- All tests pass.
- No open critical defects.
- No open high-severity defects.
- No broken user journeys.
--------------------------------------------------
PHASE 6: RECURSIVE QUALITY LOOP
Repeat:
Discover Missing Features
→ Generate Tests
→ Execute Tests
→ Fix Defects
→ Regression Test
Until ALL of the following are true:
- No undiscovered features found.
- No failing tests.
- No critical defects.
- No high-severity defects.
- No unresolved UX issues.
- No incomplete user journeys.
After each iteration produce:
1. Coverage Summary
2. Features Tested
3. Defects Found
4. Defects Fixed
5. Remaining Risks
6. Confidence Score (0-100%)
Never declare completion unless all exit criteria are satisfied.
This "loop" automation is nuts inside of Codex.
"/goal go over every single feature in this app create a user story with expected behaviour based on the code keep a single canonical spreadsheet tracking the features status
- when done switch loop to testing every user story and documenting all errors
- when done fix every logistical error or ux error
- test every user behaviour again post fix"
Shoutout to @MatthewBerman for the heads up.
Hundreds of user stories being worked through like it's nothing.
Boris Cherny and Bun's creator live-coded on stage at Anthropic's conference. They shipped 3 production PRs. The audience watched.
"This entire thing was one prompt, and that just ran for 30 minutes."
In 32 minutes, Bun's creator walks through the setup running his repo while he sleeps.
> AI bots arguing 30 rounds on every PR
> a bot that out-commits Bun's creator
> Bun being rewritten in Rust by Claude
> the image library Claude built to beat sharp
The bot writes the code. The human merges the suggestions. None of this worked three months ago.