About the project

Inspiration

Large language models have become remarkably capable at reasoning, planning and adapting to unfamiliar problems. Yet most AI agents still rely on workflows designed entirely by humans: fixed prompts, fixed memory, fixed verification and fixed execution loops.

Proteus started with a simple question:

If frontier models are now capable of reasoning about ****how****** work should be done, why force every task through the same handcrafted agent?**

Our thesis is that the model should be able to design its own task-specific workbench, choosing the workflow, prompts, memory policy, context strategy, verification process and success criteria that best fit the problem, while a trusted runtime remains responsible for execution, evidence collection and safety.

Rather than building another terminal agent, we wanted to explore a different architecture, one where the agent itself adapts to the task.

---

What it does

Proteus is a model-led adaptive runtime for terminal environments.

When given a task, Proteus first uses an Architect to synthesise a configurable workbench tailored to that objective. Depending on the task, this workbench can define:

• execution workflow
• prompts and reasoning strategy
• memory and context policies
• available tools
• verification methodology
• success criteria
• recovery behaviour

The generated workbench is then executed by a Solver, while an independent Verifier validates the resulting system state using evidence gathered from the environment rather than simply trusting the model's response.

Instead of one fixed agent, Proteus dynamically constructs the agent required for each task while keeping execution inside a trusted, evidence-driven runtime.

---

How we built it

Proteus evolved from months of research into adaptive terminal agents and was substantially extended during OpenAI Build Week using Codex and GPT-5.6.

Throughout development, Codex acted as an engineering partner for implementation, refactoring, testing and debugging, while the overall architecture, design decisions and evaluation strategy remained human-directed.

The project combines several cooperating components:

• an Architect responsible for synthesising task-specific workbenches
• a Solver responsible for execution
• an independent Verifier responsible for validation
• deterministic evaluation and certification
• evidence collection and provenance tracking
• configurable provider integrations
• evaluation boards for measuring agent behaviour and reliability

Instead of treating prompts as static configuration, Proteus treats them as artefacts generated specifically for each task.

---

Challenges we ran into

The biggest challenge was not getting an LLM to execute commands, it was building an adaptive system that could remain reliable while allowing the model to control much of its own behaviour.

Some of the problems we had to solve included:

• defining clear boundaries between model-generated policy and trusted runtime behaviour
• ensuring verification remained independent from execution
• handling provider differences and structured model outputs
• building deterministic tests for inherently probabilistic systems
• preventing false success through evidence-based verification
• balancing flexibility with reproducibility

As the project grew, packaging a research prototype into something that could be evaluated and reproduced also became a significant engineering challenge.

---

Accomplishments that we're proud of

We're proud that Proteus is built around a clear architectural idea rather than a collection of features.

Highlights include:

• a model-led configurable workbench architecture
• explicit Architect, Solver and Verifier separation
• evidence-driven execution and independent verification
• deterministic certification and evaluation infrastructure
• configurable runtime suitable for a wide range of terminal-based tasks
• extensive development using Codex throughout Build Week

Most importantly, Proteus demonstrates that agent architectures themselves can become adaptive instead of remaining permanently hardcoded.

---

What we learned

Building Proteus changed how we think about AI agents.

Initially we focused on improving prompts and execution strategies. Over time we realised the more interesting problem was deciding who designs the workflow itself.

Our biggest takeaway is that frontier models are increasingly capable of designing task-specific operating procedures, but they still benefit from a trusted runtime that enforces execution, evidence and verification.

We also learned that evaluation infrastructure is just as important as agent capability. Reliable agents require reproducible measurements, independent verification and clear evidence of success.

---

What's next for Proteus

Proteus is an ongoing research project.

Our next goals include:

• improving automatic workbench generation for increasingly complex tasks
• expanding support for additional models and providers
• strengthening memory and long-horizon planning
• extending evaluation on challenging terminal benchmarks
• enabling richer collaboration between specialised model-generated roles
• exploring self-improving workbench synthesis based on previous successful executions
• Terminal Bench 2.1 benchmark attempt with 5.4 mini and 5.6 luna. (small models due to limited budget)\

Our long-term vision is a general-purpose adaptive runtime where models do not simply execute workflows, they design the right workflow for the task while operating within a trustworthy execution environment. And those workflows become configured workbenches, like skills for harnesses.

Built With

• 5.6
• codex
• gpt