ACS · Recursive Delegation · AWPKG

Turn a question into a repeatable workflow. No code required.

Ask something. CorvinOS builds the analysis plan, runs it in the background across your data sources, and returns the findings as files you can open, share, or schedule — while the conversation stays uninterrupted.

A system that keeps improving until the answer is good enough

Instead of producing one answer and stopping, the Compute layer keeps running until the result meets your standard. It divides the problem, assigns pieces to parallel workers, checks every result, and tries again when it falls short — automatically.

TASK ROOT ★ MANAGER claude-sonnet-4-6 ITER 1 loss: 0.82 DELEGATE W-A claude-haiku-4-5 ★ SUB-MGR delegates further SW-B1 hermes SW-B2 hermes W-C claude-haiku-4-5 ITER 2 loss: 0.51 DELEGATE W-D claude-haiku-4-5 W-E claude-haiku-4-5 W-F ⚒ forge.create_tool() ⚒ FORGE TOOL created mid-run · available to all workers ITER 3 loss: 0.12 COMPLETE W-G claude-haiku-4-5 W-H ✦ skill.create() SKILLFORGE SKILL injected next turn ✓ COMPLETE quality: 93% · 3 iter · 8 workers Loss convergence 1.0 0.8 0.6 0.4 0.2 0.0 Iter 1 Iter 2 Iter 3 0.82 0.51 0.12 high loss partial converged ITERATIONS WORKERS

Problems that branch can go deeper

When a sub-problem turns out to be more complex than expected, the system spawns its own mini-team to handle it. Each branch works independently, within a shrinking budget, so the whole run stays bounded and predictable.

Every result is checked before it counts

Before a finding is accepted, it passes five independent checks: Is the deliverable complete? Is it original? Does it hold up to critique? Does it meet your declared metrics? Only results that clear all five move forward. Failed checks trigger repair, not failure.

More effort where it's needed, less where it isn't

Far from the target, the system deploys more workers in parallel to explore faster. Close to the target, it focuses on refinement. This concentrates your compute budget on the parts of the problem that are still unsolved.

Sensitive data never leaves your machine

For data classified as CONFIDENTIAL, the system automatically switches to a fully local AI — no cloud, no egress. Ordinary analysis uses cloud models. The routing happens silently, enforced at every spawn, based on your data classification.

The run teaches itself new tools as it goes

If a worker discovers mid-analysis that it needs a capability that doesn't exist yet, it creates the tool and uses it immediately — no restart, no human in the loop. When the run finishes, the custom tools are bundled into the package, ready on any system that installs it.

Ask once. Get a complete, exportable analysis.

Describe what you want to find out. The system builds the entire plan, executes it across your data, and delivers the results as files you can open, share, or schedule — while you're free to do something else.

  • 1
    Describe the goal "Analyse sales data from the last 90 days for anomalies and surface the top five outliers."
  • 2
    The system builds the plan The steps are assembled automatically — data retrieval, analysis, ranking, visualization — without any manual configuration from you.
  • 3
    Everything runs in the background The AI continues to talk with you while the compute engine works separately, applying the right strategy for the problem shape.
  • 4
    Results arrive as files Charts, tables, reports — registered automatically. The AI picks them up and can explain, summarize, or go deeper in the next turn.
  • 5
    Package it for next time If the analysis is worth repeating, export it. One package captures the entire workflow — any CorvinOS instance can run the same analysis, same steps, same quality checks, indefinitely.
YOU Describe a goal ATF Builds the task graph ACS Runs in the background RESULTS CSV · PNG · JSON .awpkg port‑ able
One coordinator. Many workers. Every result checked.
WORKERS — RUN IN PARALLEL WORKER 1 WORKER 2 WORKER 3 COORDINATOR reviews each result decides what comes next RESULTS summaries only — nothing private passes through

The coordinator reads the goal, dispatches work to parallel workers, and reviews results — but never touches your data directly. Workers run in isolation from each other and from the coordinator's context. Only clean summaries flow back up. After each round, the coordinator decides: push further, declare success, or stop cleanly.

Your data stays where it lives. The AI never sees it.

Connecting to a database does not mean your data is visible to the AI. It receives only a reference code and a structure description with personal details removed. The actual records stay behind the firewall, touched only by isolated workers that cannot communicate with the AI directly. Database passwords and keys are never in a prompt, a log, or an audit record.

LLM / AGENT SEES DATA STAYS HERE OPAQUE HANDLE a7f3b2c1d9e04f81a3b2 22 characters PII-REDACTED SCHEMA SAMPLE id: int, region: str, amount: float [email: REDACTED] [name: REDACTED] 4,000 token cap DATA FIREWALL PostgreSQL raw data ENV-INJECTED INTO BWRAP PGHOST=... · PGUSER=*** Worker 1 Worker 2 ARTIFACTS CSV PNG JSON metadata only metadata only — no content

The AI sees a reference, not your records

Instead of your data, the AI receives an opaque code. The actual records are accessible only to isolated workers — not to the conversation, not to any log.

Passwords and keys never appear in any log

Credentials are passed directly into the sandboxed worker at startup — never written to a file, never included in a prompt, never stored anywhere a later process could read them.

Sensitive data is withheld, not exposed

If the data is marked sensitive and the right key is absent, the system replaces the content with a placeholder. The run pauses cleanly — it never risks exposing confidential data to recover from the situation.

Results improve automatically until they're good enough

After each attempt, the system measures the quality of the result against five dimensions. The coordinator reviews the scores and decides whether to continue improving, accept the result, stop because progress has stalled, or abort because quality degraded. The process is self-correcting rather than single-shot.

Completeness
Everything you asked for is there — no missing pieces silently skipped.
Novelty
Each iteration is meaningfully different from the last — no treadmilling.
Quality
An independent AI critique checks whether the result is genuinely good.
Metrics
The goals you declared are measured — not assumed.
Confidence
The system's own uncertainty is factored in as a lightweight final check.
converged

Target reached, result accepted.

continue

Not there yet — another round starts automatically.

plateau

Progress stalled — the system stops rather than spinning.

regression

Quality dropped — the run stops to protect what was already produced.

The further the result is from your target, the more workers are deployed in the next round. As quality improves, the system narrows its focus. Early exploration is wide and fast; final refinement is precise and efficient.
The run can create capabilities it didn't start with

If a worker realizes mid-analysis that it needs something that doesn't exist yet — a custom formatter, a domain-specific calculator, a data normalizer — it creates the tool and uses it immediately. No human is required to intervene. The run continues, now more capable than when it started.

Any tool created this way can also be used as the target for optimization — the system will tune parameters until it finds the combination that makes the tool perform best.

New capabilities, without stopping

Workers can create tools on the fly. As soon as a tool is created, it's available to every subsequent worker in the same run — no restart, no wait.

Custom tools travel with the workflow

When you export the run as a package, any tools created during it come along. Install the package on any CorvinOS system and the tools are immediately available.

Use any tool as an optimization target

Tell the system to maximize or minimize what your tool produces. It will run hundreds of parameter combinations, evaluate each one, and converge on the best — automatically.

When the analysis succeeds, the system saves the entire run as a portable package — ready to share, install, and repeat.

From one-off question to repeatable process

A successful run doesn't disappear after you close the conversation. It becomes a package — everything needed to reproduce the same analysis, on any system, with a single install command.

ACS RUN exploratory · iterative self-tooling GRAPH BUILT DAG auto-extracted quality verified EXPORT .AWPKG manifest + workflow + provenance INSTALL awpkg install discovered-a7f3b2c1.awpkg on any CorvinOS RUN workflow run com.corvin.discovered-a7f3b2c1 deterministically
discovered-a7f3b2c1.awpkg
├── manifest.yaml              # id, version, permissions, acs_provenance
├── workflows/
│   └── discovered-a7f3b2c1.awp.yaml   # the AWP workflow DAG
└── provenance/
    ├── acs_manifest.json      # run_id, iterations, workers_spawned, duration
    ├── gate_results.json      # quality gate evaluations per iteration
    └── quality.json           # aggregate_score, evaluation count

No user data. No raw database content. Only the workflow structure and run metadata.

Fixed steps — same result every time

The workflow is frozen exactly as it ran. Deploy it on a schedule and you get the same analysis, the same way, indefinitely.

awpkg export --mode dag

Guided structure — adapts to new data

The shape of the workflow is fixed, but the execution adapts. Useful when the underlying data changes and you want the analysis to stay current without rebuilding from scratch.

awpkg export --mode template

Every result is traceable

The package includes a full record of how the result was produced — iterations, quality scores, every gate evaluation. You can always trace a finding back to exactly how it was generated. Apache-2.0 license is applied automatically.

Install
# Any CorvinOS instance
awpkg install discovered-a7f3b2c1.awpkg

# Or from a URL / shared path
awpkg install https://share.example.com/discovered-a7f3b2c1.awpkg
Run
# Deterministic execution
workflow run com.corvin.discovered-a7f3b2c1

# List installed packages
awpkg list

# Inspect before running
awpkg inspect discovered-a7f3b2c1.awpkg
Start with a question. End with a scheduled process.

Discovery and repeatability are two different needs. Use the exploratory mode when you don't know the steps yet, switch to the deterministic mode once you have a proven workflow you want to run on autopilot.

Exploratory

ATF — AI-generated workflow

You describe what you want to find. The system figures out the steps, tries different approaches, and adapts as results come in — you don't need to know the method in advance.

orchestration.engine: delegation_loop
  • Start with a question — the steps emerge as the analysis runs
  • The right search strategy is chosen automatically for your problem
  • The analysis adjusts itself based on what it finds along the way
  • Results arrive as files — charts, tables, reports
  • Turn a successful run into a step inside a larger scheduled process
Deterministic

AWP DAG — fixed steps, scheduled runs

Once you know the right process, run it on autopilot. Every step is defined, the path is fixed, and the results are identical each time — on a schedule, or triggered by an event.

orchestration.engine: dag
  • Built for processes you trust and want to run without supervision
  • Same input, same output — guaranteed
  • Schedule it daily, trigger it from a webhook, or chain it to another agent's result
  • Turn last week's exploratory run into today's automated workflow
  • Or design the pipeline yourself step by step
Explore once. Install anywhere. Run forever.

When an exploratory run produces something worth repeating, you don't have to rebuild it. Package it, share it with a colleague or a second deployment, and run it again — no re-configuration, no manual steps.

🔍

Exploratory run

Describe the goal. The system builds the plan, runs the analysis, and delivers findings as files.

📦

Export as .awpkg

One command captures the entire workflow — steps, tools, and quality checks — as a portable file.

🔁

Share & install

Send the package to a colleague or deploy it on another system. One install command is all it takes.

🕐

Run deterministically

The same analysis, on a schedule, with the same quality guarantees — indefinitely.

What an AWPKG contains
AWP workflow definition
The complete step-by-step plan of the run, preserved exactly as executed
Forge tools
Custom tools created during the run, bundled and ready to use on the next system
Skills
Behavioral instructions refined during the run, injected automatically on execution
Manifest
Everything needed to reproduce the run — except the data itself, which stays where it lives
Your data never needs to leave your infrastructure

Every compute worker runs on your own machines — on-premises, your own cloud VMs, or wherever your databases already live. CorvinOS never requires data to travel to a third-party service.

🏢

Deploy where your data already lives

Workers run on your own servers or cloud VMs. No third-party infrastructure is required to process your data.

Scale by adding workers, not configuration

Add more worker instances to handle bigger workloads. Jobs are distributed automatically — no manual scheduling.

🔒

Data stays inside your network

Analysis runs against data sources inside your perimeter. Results flow back to the conversation — the raw data never leaves.

📦

Workflows move freely between systems

Packages install on any CorvinOS deployment. Move a workflow between environments without rebuilding or reconfiguring.

🎯

Each team's work is fully isolated

Worker queues, results, and artifact stores are sandboxed per organization. One team's analysis cannot touch another's — by design, not policy.

🇺🇪

Data residency enforced at every step

Each compute spawn is checked against your data classification and network rules before it runs. If the check fails, the job stops — it never proceeds on a guess.

The right search strategy, chosen automatically

For most questions, you don't choose a strategy — the system picks for you. But if you know the shape of your problem, you can pin one. All three can run as part of a larger pipeline.

Grid Search

Checks every possible combination. Use this when missing an answer is not an option and you need complete coverage of the parameter space.

Choose this when completeness matters more than time

Random Search

Samples the space broadly without trying every point. Reaches good results quickly, even in complex spaces with many variables.

Choose this when the space is large and you need answers fast

Bayesian Optimization

Learns from each attempt and focuses subsequent effort on the most promising directions. Reaches good results with far fewer total evaluations.

Choose this when every evaluation has a cost and you need to minimize the number of attempts
Every step is recorded and tamper-proof

Every compute run, every result, every failure is written to a log that cannot be altered retroactively. You can trace any finding back to the exact run that produced it — and prove nothing was modified after the fact.

Tamper-proof audit log
Records can't be altered or deleted without detection — the chain breaks if anything is modified.
Tenant isolation
One organization's compute jobs, results, and records are invisible to every other.
No data in the log
The log records what happened and when — never the content of your data or your prompts.
Gate before every job
Before any job starts, the system checks that the data is allowed to reach the chosen engine. No exceptions, no manual overrides.

Start with a question. End with a workflow you can schedule.

CorvinOS is open source. Install it, describe what you want to find out, and the compute layer handles the rest — no account, no vendor dependency.

View on GitHub →