Agent context architecture
Workspaces laid out so an agent can read the whole operation: one entry point per system, persistent memory, and every decision recorded where the next run will look for it.
RunningSheet 01Control lineRev 2026-09
Instruments an agent can operate. Systems that run unattended.
Darko AI is an experimental technology company. One agent layer operates every system on this sheet: a media studio, distribution, commerce, the plant that keeps it safe, and the research instruments now in build. Follow the line.
ST-01
Agent layer
One agent layer operates everything below it. It reads the operation as a map, not as a prompt.
Run ledger and handoff protocol
Every session ends in a verifiable handoff: state, repeatable checks, known traps. The next run starts from evidence, never from memory.
RunningST-02
Media production
Research, script, voice, footage, edit and upload, unattended, on fixed daily slots.
Long-form documentary studio
A daily episode is researched, scripted, voiced, cut and uploaded without a person in the loop. It cannot publish until it passes gates for pronunciation, footage-to-claim match and measured audio quality.
RunningShort-form vertical clip line
Headline to published clip in one pass. Footage is checked against the story's current state, and any shot that contradicts the claim is a hard block.
RunningDigital publishing engine
Books written, illustrated, typeset and listed as a pipeline. Interior art is applied and the whole book rebuilt on every change, so nothing stale reaches a reader.
RunningSynthetic presenter system
A consistent on-screen persona with a locked identity, upscaled locally at no per-frame cost, used across creative and reception footage.
RunningPerformance creative engine
Ad concepts are generated in families, scored, then gated for variety so no two shipped concepts share a hook.
RunningSynthetic voice studio
Reference-engineered narration with matched-window audio forensics. Quality is measured on the waveform, never asserted.
RunningGenerative media experiments
ParkedMusic and image lanes, parked until the cost card clears.
ST-03
Distribution and growth
Where the output goes, and the guards that stop it going twice.
Search and discovery automation
Keywords harvested from live suggestion surfaces, pages generated to fill the gap, then re-ranked on what the engine actually returns.
RunningAutonomous social operator
A public-facing agent that reads its timeline, replies in character and posts on its own cadence. Running since September 2026.
RunningSocial distribution mesh
Scheduled publishing across the major networks plus a federated server we host and operate ourselves.
RunningPartner outreach lane
Deliverability-checked email sequences with cooldown guards, so a partner is never contacted twice in one window.
RunningContinuous broadcast channel
A 24-hour ambient stream assembled from generated audio and footage. In build.
In buildCold outreach platform
ParkedSelf-hosted sending infrastructure, parked pending a vendor settlement.
ST-04
Commerce
Listings that cannot ship half-built.
Marketplace publishing engine
Product pages, art, descriptors and delivery files generated and published across two marketplaces. Slug locks and footer gates hold a listing back until every part exists.
RunningUtility web network and promo engine
ParkedSingle-purpose tools and their promotional pages, parked until distribution justifies them.
Reputation defence system
ParkedMonitoring plus an AI voice agent that answers on a client's behalf, parked awaiting a market signal.
ST-05
Infrastructure and safety
Idle compute is a fault. An unguarded endpoint is a fault. An unmeasured model chain is a fault.
Compute governance
A reaper polls two GPU markets every minute, kills anything older than its budget and enforces a hard daily spend cap.
RunningEndpoint sentinels
A camera presence guard on the workstation, key-only hardened servers with intrusion lockout, and a local audit daemon logging every privileged action.
RunningResearch infrastructure
Self-hosted metasearch, a fetch relay on a separate network path and a research API, so nothing upstream can rate-limit a pipeline.
RunningModel routing bench
Every model chain is re-benchmarked on its own prompt, three repeats minimum, and ranked on failure mode before it ships.
RunningST-06
Experimental hardware
Research instruments an AI can discover, plan against and operate, with every limit enforced below the software. Nine programs, one platform, one loop.
M01 In sourcing
Programmable optical stimulation platform
A modular bench instrument with independently addressable optical channels across a defined optical band, each with its own drive current, pulse frequency, duty cycle and timing. Engineering-partner RFQs are out; the controller, module bus and safety envelope are specified.
M02 Specified
AI-operable instrument layer
REST, USB serial and BLE plus a native MCP server, so coding agents discover and operate the instrument through documented tools instead of driving pins.
M03 Specified
Self-describing hardware
Capability discovery and a digital twin: one machine-readable description of modules, sensors, limits, calibration history and firmware.
M04 Specified
Deterministic protocol engine
Experiments are machine-readable protocols executed by a deterministic engine behind a safety validator. An agent proposes; hardware limits decide.
M05 Specified
Calibrated sensing and closed loop
Irradiance, current, temperature, distance and ambient light measured and logged; measured output fed back to hold a calibrated setpoint.
M06 Specified
Instrument synchronization
Hardware trigger I/O and a monotonic clock so spectrometers, cameras and external sensors share one timeline with every optical event.
M07 In design
Circadian light research module
A separate photic module for timed, low-level light with its own far more conservative hardware envelope, kept out of the free-form optical engine.
M08 Research program
Cellular energy research vectors
Mitochondrial state, redox signalling and cellular clearance treated as measurable variables alongside timing, dose and non-device interventions.
M09 Building
The experiment loop
Hypothesis to protocol to safety validation to experiment to telemetry to analysis to the next hypothesis. Agents design and analyse; the deterministic layer runs.
Detail A
The control chain
How an agent reaches the hardware. Every request crosses a validator that reads limits from the device itself; the model never holds the keys.
Supply
Built for clients
The same crew, the same gates, on your operation.
Autonomous agents
Agents that own an outcome, not a prompt
A workflow waits to be triggered. An agent decides. We build systems that run a business function on their own schedule: they find the work, research it, choose an action, take it, and follow up without anyone opening a dashboard.
OpenGrok-native bots
Bots that live inside X, at timeline speed
Most social automation is a scheduler with a text box. A Grok-native bot reads the live timeline, understands what is actually happening right now, and answers inside the window where a reply still earns reach.
OpenCustom builds
The system that does not exist yet
The work that has no product to buy. Private model routing, browser automation fleets, render pipelines, ingestion and enrichment, internal tools that three people use forty times a day.
OpenNotes
Read before operating
- 01
Safety is enforced in hardware
Exposure timers, current limits, thermal cut-offs and watchdogs live in silicon and firmware. A computer, a model or a prompt is never the last line of defence.
- 02
The instrument describes itself
Every device publishes its modules, sensors, limits, calibration and supported commands as a machine-readable schema. Software discovers hardware; it does not assume it.
- 03
Model-agnostic by design
Hardware and protocol engines run without any AI model. Claude, OpenAI, local models and plain scripts all talk to the same documented API, USB, BLE and MCP surfaces.
- 04
Measured, not assumed
Optical output, temperature and timing are measured with calibrated sensors and logged with synchronized timestamps. Nominal specifications are inputs, not evidence.
End of line
Bring the system you cannot run by hand.
Thirty minutes. We map it onto this sheet, name the gates it needs, and tell you whether it is an agent build, an instrument, or both.