ChatCortex
ChatCortex is a framework for automated synthesis and optimization of AI agent architectures — think AutoML for AI agents.
Instead of manually wiring LLM pipelines (retrieval → LLM → verifier → tool), ChatCortex treats agent design as a multi-objective architecture search problem and automatically discovers architectures that optimize cost, latency, and reliability simultaneously.
Task Specification → Architecture Search → Pareto-Optimal Architectures
ChatCortex re-synthesizes your architecture from what was measured. Execution telemetry is calibrated back into the component registry, so every new synthesis round runs against observed behavior instead of declared guesses. examples/closed_loop.py tests this offline against a hidden ground truth: the starting registry is wrong about four components, and the simulator draws outcomes from true values the loop never sees. After sixteen iterations with evidence-weighted calibration (prior_strength=20, seed 42, 677 probes):
| Component | Field | Declared | Calibrated | Truth |
|---|---|---|---|---|
retrieval_speedster |
reliability | 0.78 | 0.707 | 0.55 |
retrieval_budget |
reliability | 0.82 | 0.945 | 0.95 |
generation_balanced |
latency (ms) | 180 | 281.1 | 320 |
verification_reliable |
cost per call | 0.006 | 0.0080 | 0.009 |
One seed is one draw, so the example also repeats the run over 20 seeds and reports mean reliability error across all twelve components:
| Calibration rule | Mean reliability error | Runs that beat the declared registry |
|---|---|---|
| None (declared values) | 0.030 | n/a |
Fixed alpha=0.4 |
0.066 | 0 of 20 |
Evidence-weighted, prior_strength=20 |
0.024 | 17 of 20 |
What this does and does not show. Evidence-weighted calibration improves on the declared registry given enough probes; a fixed blend weight makes reliability worse. It needs the evidence: in the same test at four iterations it does not yet beat the declared values on average. A component that drops off the frontier stops being probed, which is why
retrieval_speedsterabove is still far from its true value.
Installation
pip install chatcortex
Or from source:
git clone https://github.com/siddharth1012/chatCortex
cd chatCortex
pip install -e .
Quickstart
Want to know which free model and prompt answer your own questions best? Follow the guide: a CSV of questions, a few free API keys, one command.
from chatcortex import (
TaskSpecification,
BeamSynthesizer,
build_demo_registry,
)
# A registry of components with cost/latency/reliability tradeoffs
registry = build_demo_registry(
capabilities=("retrieval", "generation", "verification")
)
# The task: an ordered capability chain with hard constraints
task = TaskSpecification(
required_capabilities=["retrieval", "generation", "verification"],
max_cost=0.02,
max_latency=1500,
)
# Synthesize Pareto-optimal architectures
synth = BeamSynthesizer(registry, beam_width=5)
for arch in synth.synthesize(task):
print(
f"cost=${arch.total_cost:.4f} "
f"latency={arch.total_latency:.0f}ms "
f"reliability={arch.total_reliability:.3f}"
)
Or run the full demo comparing three synthesis strategies:
python examples/quickstart.py
To see the full loop — load a JSON registry, synthesize a frontier, execute one architecture, and calibrate the registry's metadata from real telemetry — run:
python examples/registry_from_json.py
Run the flagship closed loop — synthesize → execute → calibrate → re-synthesize, four iterations of measured behavior flowing back into the registry:
python examples/closed_loop.py
With an OpenAI key you can run the same loop against real backends:
pip install "chatcortex[real]"
export OPENAI_API_KEY=... # every reply spends your credits
python examples/closed_loop_real.py
Give the backend your prices and each call is costed from the token usage the
API reports, instead of the declared cost_per_call:
from chatcortex.adapters.openai_backend import OpenAIBackend, TokenPrice
backend = OpenAIBackend(
model="your-model",
prices={"your-model": TokenPrice(input_per_million=2.0, output_per_million=8.0)},
)
Prices are yours to supply because they are model-specific and change over time. A model without a price falls back to the declared cost.
Other vendors and local servers
Many model providers and local servers accept the same chat-completions
request format. Point the backend at one with base_url and api_key; no
extra dependency is needed. RoutedBackend sends each component to its own
backend, so one search can compare models from several vendors:
import os
from chatcortex import RoutedBackend
from chatcortex.adapters.openai_backend import OpenAIBackend
backend = RoutedBackend({
"writer_a": OpenAIBackend(model="model-a", base_url="https://vendor-a.example/v1",
api_key=os.environ["VENDOR_A_KEY"]),
"writer_b": OpenAIBackend(model="model-b", base_url="https://vendor-b.example/v1",
api_key=os.environ["VENDOR_B_KEY"]),
})
examples/compare_endpoints.py scores two providers on the same three
questions and prints measured quality and latency for each. It reads
SAIA_API_KEY, NVIDIA_API_KEY, GROQ_API_KEY, OPENROUTER_API_KEY and
ANTHROPIC_API_KEY from the environment and runs whichever are set; edit its PROVIDERS table to use others. Request
options are set per provider, because vendors differ in what they accept:
current Claude models, for example, reject a temperature below 1.
To search over your own components, build a CapabilityRegistry and register
ComponentMetadata describing each model, tool, or verifier — see
examples/ for complete walkthroughs.
Run your own functions
Back each component with a plain Python function and the executor runs a synthesized pipeline for real: every function receives the previous step's output and returns its own.
from chatcortex import AgentExecutor, CallableBackend
backend = CallableBackend({
"retrieve_keyword": lambda question: search(question),
"answer_short": lambda passages: draft(passages),
"verify_grounded": lambda answer: check(answer),
})
backend.missing(registry) # [] when every registry component is covered
result = AgentExecutor(backend=backend).execute(arch.graph, task_input="your question")
result.output # what the last step returned
result.success # False if any step raised or reported failure
Latency is measured around each call. A function that raises, or returns
StepResult(output=None, success=False), is recorded as a failed step and
stops the pipeline. Cost is the declared cost_per_call unless the function
returns StepResult(output, cost=...). A complete runnable version:
python examples/callable_pipeline.py
Score pipelines on your own examples
Declared numbers say nothing about whether a pipeline's answers are any good. Give ChatCortex example inputs and a scoring function, and it runs each pipeline on every example and reports measured quality, cost and latency.
from chatcortex import EvalSet, Example, evaluate_pipelines, quality_frontier
eval_set = EvalSet(
examples=[
Example("What is the boiling point of water?", expected="100"),
Example("How quickly does light travel?", expected="299,792"),
],
scorer=lambda output, expected: expected in str(output), # bool or 0..1
)
evaluations = evaluate_pipelines(frontier, executor, eval_set)
best = quality_frontier(evaluations)[0] # highest measured quality first
best.components, best.quality, best.cost, best.success_rate
A run that fails part-way scores 0. quality_frontier keeps the pipelines not
beaten on measured cost, latency and quality. In examples/scored_pipeline.py
the three frontier pipelines score 1.00, 0.00 and 0.00 on four questions,
although their declared reliabilities are 0.91, 0.85 and 0.71.
Only pipelines the synthesizer proposes are scored. Pass eval_set= to
run_closed_loop and measured quality steers the search: each component gets a
quality_score estimated from every scored run, and the next round is
synthesized with quality as a fourth objective next to cost, latency and
reliability. Components not yet measured count as 1.0, so they get tried. A
pipeline is scored once and reused in later rounds.
records = run_closed_loop(
registry, task,
synthesizer=lambda reg: ExhaustiveSynthesizer(reg, keep_ties=True),
executor=executor, eval_set=eval_set, prior_strength=20, iterations=8,
)
keep_ties=True keeps every pipeline with identical predicted numbers instead
of one of them, so look-alike alternatives (two context options with the same
declared latency, say) all get measured.
Once every pipeline on the predicted frontier has been scored, the loop stops
trying new ones, even if an untried one would be better. explore=N also scores
the N untried pipelines with the highest predicted quality each round, so it
keeps looking at the price of more calls.
python examples/scored_pipeline.py
Registry packs
A registry ships as a JSON pack in the loader's document shape. Saving it proves the round-trip, and one command validates any pack offline:
from chatcortex import build_demo_registry, save_registry, validate_registry_file
save_registry(build_demo_registry(), "registry.json") # writes, then reloads to prove it
print(validate_registry_file("registry.json").valid) # True
chatcortex-registry validate registry.json
# components validated: 12
# result: VALID (exit 0; issues and duplicates are named when invalid)
How It Works
ChatCortex is organized as a layered architecture synthesis system:
TaskSpecification → CapabilityRegistry → Synthesis Engine → AgentGraph (DAG)
→ Execution Engine → Telemetry → Evaluation Harness → Pareto Optimization
Core concepts:
- ComponentMetadata — declarative description of an agent component (model, retriever, tool, verifier, memory) with cost per call, latency, reliability score, and privacy level.
- CapabilityRegistry — stores components and filters candidates by capability and privacy constraints.
- TaskSpecification — the synthesis problem: an ordered capability chain plus hard constraints (max cost, max latency, privacy) and objective weights.
- AgentGraph — architectures represented as DAGs; cost is additive, latency sequential, reliability multiplicative.
- Synthesizers — search strategies that map a task to an (approximate) Pareto frontier.
Available synthesizers:
| Synthesizer | Strategy |
|---|---|
HeuristicSynthesizer |
Greedy deterministic construction |
RandomSynthesizer |
Random baseline |
BeamSynthesizer |
Budget-aware beam search |
ParetoPartialBeamSynthesizer |
Beam search with per-stage Pareto pruning |
ProgressiveParetoBeamSynthesizer |
Depth-aware beam widening (best Pareto recovery) |
ExhaustiveSynthesizer |
Exact Pareto frontier (small spaces) |
Approximation quality is measured with Pareto coverage, hypervolume loss, and average regret — see docs/BENCHMARKS.md.
Documentation
- Guide: pick the best models for your questions, for free — compare free models and prompts on your own questions and get a recommendation
- Architecture — system design and component layers
- Benchmarks — experimental evaluation of synthesis strategies
- Examples — runnable demos and budget-sweep experiments
- QA frontier benchmark — finds the pipelines worth choosing between on real questions, using free models only
- Contributing — development setup and PR guidelines
- Changelog
Research Context
ChatCortex is a controlled experimental platform for studying automated agent architecture synthesis: multi-objective optimization, architecture search, reliability–cost tradeoffs, and AutoML-style agent design. Experiments cover 5-stage pipelines with Pareto frontiers up to 95 architectures and evaluation budgets from 20 to 180 evaluations; progressive beam widening shows improved Pareto recovery over static beam strategies.
If you use ChatCortex in your research, please cite it (see CITATION.cff).
Roadmap
- Component modeling, capability registry, task specification, DAG representation
- Exact Pareto frontier computation (exhaustive search)
- Budget-aware beam search with Pareto pruning and progressive widening
- Graph-structured (non-chain) agent synthesis
- Real model / tool integrations (OpenAI chat-completions backend already available behind the
chatcortex[real]extra; more backends planned) - Registry loading from JSON (
load_registry, with error messages naming the failing field); YAML planned
Status
Research framework under active development (alpha). APIs may change between minor versions.
License
MIT — developed by Siddharth Saraswat
Metadata
Release files for chatcortex 0.10.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| chatcortex-0.10.1.tar.gz | 80.8 kB | With provenance | Details for chatcortex-0.10.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| chatcortex-0.10.1-py3-none-any.whl | Python 3 | none | any | Details for chatcortex-0.10.1-py3-none-any.whl |
Total release size: 141.0 kB
Release files / chatcortex-0.10.1.tar.gz
| Download URL | chatcortex-0.10.1.tar.gz |
|---|---|
| Size | 80.8 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
5a6a00ab28826e5529d9183ead4ec6f520443f53880882ad7c9df608cc27bf29
|
|
BLAKE2b-256 checksum How to use checksums |
7940cc23e29c85b35fb05b098d7494ae6d7ede4cc753f112a0258a4d44d78eeb
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.
Transparency logRelease files / chatcortex-0.10.1-py3-none-any.whl
| Download URL | chatcortex-0.10.1-py3-none-any.whl |
|---|---|
| Size | 60.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
beecda327374cbf08bea1ce2cd5dffefe3c98e19cc1baf578335fd3bbcd32201
|
|
BLAKE2b-256 checksum How to use checksums |
94cfc6fc64a1173f5a81b78e538fea2a9138820aa6efc9a06002748f89f2391e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.
Transparency logSecurity for chatcortex 0.10.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for chatcortex-0.10.1.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 19, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.