AI Agent Operations Studio · Indonesia · Founded 2026

Autonomous agents.
Human command. Verified execution.

The control layer for autonomous Web3 work.

AkTa Agent Works is an AI agent operations studio building the control layer for autonomous Web3 work: persistent agents that research, decide with specialist models, execute under explicit limits, and prove every material result.

Other agents optimize for acting. We optimize for acting only when the evidence, the limits, and the final state can be verified.

263files under a SHA-256 integrity manifest — verified 263 OK · 0 changed · 0 missing · 0 new
104/104tests passing across five core safety & operations tools
48active agent skills in our verified library
8 OK · 0 FAILlive system health check — with 1 warning disclosed, not hidden
$0raised — self-funded, by design and by record

Product

AkTa Control Stack

AI agents can already search, call tools, and act. The unsolved problem is operational trust: fragmented data, unbounded spend and permissions, silent changes to skills, and completion claims nobody can audit. Control Stack is the safety and operations layer we built to close those gaps — and the system our own agents run on every day.
Private alpha — in daily production use by our own operations

Skill Library & Integrity

48 reusable agent skills — research, contract audit, bounty operations, DeFi, LLM security, transaction pipelines — hashed into a SHA-256 manifest. Any changed, missing, or new file is detected before it can quietly alter agent behaviour.

263 files verified · 8/8 tests

Spend Governor

Fail-closed spend authorization with per-transaction, daily, and session caps. Incomplete state means deny. A manual HALT file stops all spending instantly.

34/34 tests

Cost Ledger

An append-only book of API and on-chain spend with daily thresholds and alerts — the audit trail every governor decision is checked against.

23/23 tests

Transaction Pipeline Gates

Every on-chain action follows build → simulate → screen → authorize → confirm → broadcast → record, with receipt and final-state verification before anything is called a success.

Operating rule — enforced daily

Confirmation Codes

One-time confirmation codes bound to the SHA-256 hash of the exact transaction payload. Change one byte and the code is void.

Protocol in active use

Log Redaction

Two-layer redaction for private keys, bot tokens, session strings, credentialed proxy URLs, mnemonics, and sensitive environment values — while keeping operational addresses auditable.

20/20 tests

Process Watchdog

Liveness judged by CPU-time progress, not by process-name matches that can fool themselves. Bounded restarts with a per-hour cap and an operator-cleared halt state.

19/19 tests

System Doctor

One read-only command that reports the health of the whole deployment — proxies, wallet sessions, disk, integrity manifest, governor state, and system clock.

Live run: 8 OK · 1 WARN · 0 FAIL

Shared Session Memory

An append-only shared memory (SQLite + full-text search) storing decisions and stop-rules across sessions and between our two agent operators — with secrets refused at write time.

In daily use since 20 Nov 2026

Alongside the stack, the AkTa Operations Desk is the service that runs on it: project and airdrop research, agent-competition operations, on-chain operations, testnet market-making research, prediction-market evidence work, and bug-bounty research — every workflow ending in a verifiable artifact, a decision, or a gate that stopped the action.

How it works

Every task runs the same four steps

No step is skipped because an agent "feels confident". The loop is the product.

STEP 1

Research

Evidence is gathered from primary sources — social signals, search, market data, source pages — and assembled into a dossier with provenance: source, read time, and whether each value was observed, reported, or estimated.

STEP 2

Specialist reasoning

Frontier and specialist models reason over the dossier in defined seats — strategy, critique, adjudication. The operator carries their findings back faithfully; it does not replace specialist thinking with its own summary.

STEP 3

Gated execution

Actions pass simulation, preview screening, spend caps, and explicit human approval for anything valuable, public, risky, or irreversible. A gate that fails stops the action — that is a success state, not an error.

STEP 4

Verification

Nothing is called done from a tool's output alone. Receipts, final on-chain state, files, and live readings are checked before a result is reported — and the evidence travels with the report.

Evidence

Traction, stated plainly

Our traction today is product and operations traction, not market traction. We publish the numbers with their context — including the ones that are not revenue.

Disclosed warning: our live System Doctor run reports 8 OK, 1 WARN, 0 FAIL. The warning is real: the Spend Governor is installed and fully tested, but has not yet been wired into a production runner. We would rather show the warning than hide it — that is the product.

Team

One human. Two persistent agents.

We describe the team exactly as it is: 3 core contributors — 1 human founder/operator and 2 persistent AI agent operators. That structure is not a novelty; it is the operating model our product exists to make safe and auditable.

S

Sara

Founder & Operator-1 — Human

Sets direction, portfolio priorities, and risk limits. Holds final approval for every valuable, public, risky, or irreversible action. Based in Indonesia.

A

AkTa

AI Agent Operator & Orchestrator

Coordinates research, specialist models, tools, schedules, execution, and verification across the project portfolio. Builds and maintains the Control Stack.

S

Salmia

AI Agent Operator & Research Counterpart

An independent second operator: research counterweight, a shared rulebook for joint competition work, and structured exchange of operational memory. Runs its own separate agent account.

Why AIsa data

Data is our production input

Before an agent recommends, builds, submits, or executes anything, it needs current evidence — today fragmented across social platforms, search, market sources, and web pages. AIsa's catalog, with provenance on every value, matches our operating discipline: no material claim without a source, no execution without verification. Our planned 12-month allocation of the $10,000 Build-tier balance:

CategoryWorkflow it powersAllocation
XOpportunity radar: project announcements, launch signals, competitor & agent-cohort monitoring$1,600
Crypto market dataProject triage: token, price, liquidity, and market context before any decision$1,500
Deep ResearchStructured evidence dossiers for triage, competitor cohorts, and bug-bounty scope research$1,500
Google SearchDiscovery, official documentation, news, and claim verification$1,200
Page ExtractionClean evidence from project docs, whitepapers, bounties, changelogs, and dashboards$1,100
TikTokVideo, creator, and trend intelligence for clipping and growth competitions$900
YouTubeVideo and channel research for clipping, launches, tutorials, and competitor analysis$900
Prediction MarketsEvidence mining and market context for prediction-market workflows$800
InstagramProject and creator checks, visual campaigns, impersonation monitoring$500
Total — 12 months. Every call tied to a workflow, recorded in our cost ledger, and expected to produce an evidence file, a decision, or a verified result.$10,000