Just checked my Base network score, didn’t expect 10000 RA points free on @remoteaixyz 👀
Check yours too, this might be big — theremoteai.xyz/check-score?re…
English
CryptoPlug ↑↑↑
1.5K posts
CryptoPlug ↑↑↑
@CryptoOGPlug
$CRED & $OSO Maxi, $CRED & $OSO Believer. $WHITE
London, England Tham gia Mayıs 2023
371 Đang theo dõi1.1K Người theo dõi
Just checked my Base network score, didn’t expect 10000 RA points free on @remoteaixyz 👀
Check yours too, this might be big — theremoteai.xyz/check-score?re…
English
CryptoPlug ↑↑↑ đã retweet
Remote AI is building autonomous agents to optimize token launches by bridging liquidity markets in real time.
Exclusively built on @base
Your past Base network activity now earns RA points — redeemable for $RA 👀
Check yours: theremoteai.xyz
English
CryptoPlug ↑↑↑ đã retweet
agenc one os just got a major update.
switched to xai's latest tts voice api and honestly the voice quality is insane, this thing sounds natural, fast, no robotic artifacts.
>running grok 4.20 for the brain 👾
agencone.com
English
CryptoPlug ↑↑↑ đã retweet
big AgenC //: 👾 runtime push landed today
agents now run as background processes. they survive daemon restarts, resume with full task context, and reattach to the same live desktop workload.
durable state, boot-time recovery, process supervision, native restart paths.
ran it through runtime, desktop-server, and web tests. did live restart and resume runs.
long-running agents that can't survive a restart aren't useful.
English
CryptoPlug ↑↑↑ đã retweet
👾 AgenC is preparing for external audit.
Focus:
- freeze critical scope.
- strengthen test/fuzz coverage.
- close remaining Skill Marketplace integration gaps.
- finalize threat model + audit package.
- prepare external audit.
NO mainnet claims before audit. We’ll share progress publicly
github.com/tetsuo-ai/Agen…
English
CryptoPlug ↑↑↑ đã retweet
A ZK proof lets you prove something is true without revealing the private details behind it.
Simple example:
You can prove you solved a task correctly, without publishing your full raw output.
Why this matters in AgenC:
- Work can be verified
- Sensitive output can stay private
- Payments can still be released by objective rules
Think of it like:
“Here is cryptographic evidence I did it right”
instead of
“Trust me, I did it.”
This is how AgenC combines verifiability + privacy in one flow. 👾
English
CryptoPlug ↑↑↑ đã retweet
Let’s make this simple.
When a task is created, payment is locked first in a smart contract. That locked payment is called escrow.
Think of it like this:
A trusted locker holds the money until the work is completed and verified.
Why it matters:
Worker knows funds are reserved
Creator knows funds are released only by rules
No need to “trust me bro” between strangers
In AgenC, escrow is the trust layer that makes task coordination safer before we even talk about privacy or ZK proofs.
English
CryptoPlug ↑↑↑ đã retweet
Official Notice - Tetsuo Corp. / AgenC
As of March 4, 2026, our official channels are only:
Code: github.com/tetsuo-ai/AgenC
Website: agenc.tech
Website: agencone.com
Docs: docs.agenc.tech
We currently have no official npm package, no official mobile app, and no official browser extension published outside these channels.
Any package/app claiming to be official outside the links above is unauthorized and should not be trusted.
CA: 5yC9BM8KUsJTPbWPLfA2N8qH1s9V8DQ3Vcw1G6Jdpump
English
CryptoPlug ↑↑↑ đã retweet
Something very important we need to keep in mind about agenc.tech is that its portability and open source architecture give it a major advantage for lunar exploration. Because it can run across different hardware environments and adapt to multiple computing platforms, AgenC can be deployed on rovers, landers, orbital modules, or ground-control systems without heavy redesign. Its open source nature allows seamless integration with a wide range of sensors and scientific instruments, enabling real time data processing, autonomous decision making, and flexible upgrades as mission needs evolve. In a high risk, resource constrained environment like the Moon, this adaptability means faster iteration, lower integration costs, and stronger collaboration between international teams and research institutions.
@tetsuoai
English
CryptoPlug ↑↑↑ đã retweet
AgenC ONE update: the agent framework is now an OS running natively on a Pi Zero, controlling its own display, sensors and Solana wallet
AGENC@a_g_e_n_c
we finally did it! agenc-one now runs its own linux distro built from scratch. not ubuntu, not raspbian, nothing pre made. our own OS, our own kernel, our own partition layout custom read only rootfs so nothing breaks. writable data partition for the agent state. auto generated SSH keys on first boot. firewall locked down, only outbound allowed 4 partitions on a tiny sd card. boot, two rootfs slots for OTA updates, and a data partition that survives everything the pi boots straight into our runtime. no desktop, no bloat, no gui. just the agent full control of every single byte running on the hardware this is what building from zero actually looks like
English
CryptoPlug ↑↑↑ đã retweet
Here is a live view of @tetsuoai coding for $AgenC right now 🫡
GIF
TΣTSUΩ@tetsuoarena
my day: find vulnerabilities. Patch them. run tests. Update tests. Write new tests. run them again. boot the OS on the Pi. Analyze performance. New build. Test new features. Repeat. create content for the community. Repeat. also apparently forgetting basic human needs(🛀)
English
CryptoPlug ↑↑↑ đã retweet
Agenc team never rests, we just shipped SOTA subagent orchestration and rebuilt the entire terminal from the ground up.
19k+ lines, learning-based delegation, on-chain coordination via Solana v0.2 is live.
tetsuo@tetsuoai
╔═══════════════════════════ ║ A G E N C v 0 . 2 ║ ── ORCHESTRATION + TERMINAL ── ╚═══════════════════════════ we built SOTA subagent orchestration. the agent decomposes tasks into DAGs, spawns child agents with scoped tools and curated context, verifies their outputs, synthesizes results with provenance, and learns from outcomes to improve policy over time. 9 workstreams. 108 files. ~20k lines. grounded in primary research: AOrchestra dynamic agent tuples BOAD bandit optimization Agent Lightning trajectory logs AFlow DAG workflow search Magentic-One manager baseline SWE-search tree decomposition Anthropic pass@k vs pass^k OpenAI multi-agent SDK ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ┤ ORCHESTRATION ENGINE ├ SubAgentOrchestrator materializes planner DAGs into executable work. serial and bounded-parallel modes. built on DeterministicPipeline- Executor. preserves completed / failed / halted + resume semantics. ChatExecutor planner now emits: - deterministic_tool steps - subagent_task steps - synthesis steps each subtask specifies: objective, input_contract, acceptance_criteria, required_tool_capabilities, context_requirements, max_budget_hint, can_run_parallel ┤ DELEGATION INTELLIGENCE ├ DelegationDecision scorer computes: - decomposition benefit - coordination overhead - latency/cost risk - safety risk - confidence score gated by threshold + guardrails. explicit "do not delegate" path for trivial requests. robust validation: rejects malformed graphs prevents cycles enforces max fanout + depth structured diagnostics ┤ CONTEXT CURATION ├ AOrchestra-style child packing: - targeted history slice - relevant tool outputs only - required memory results only - minimal prompt envelope no full parent history dumps. context size caps + diagnostics. least-privilege tool scoping: child allowlist from parent policy + subtask requirements. privilege expansion blocked. deny-list and SSRF preserved. ┤ VERIFIER + SYNTHESIS ├ verifier stage for child outputs: - contract adherence - evidence quality - hallucination detection - tool-result consistency low-confidence rejected or retried. parent synthesis includes: child outputs + confidence failure markers + unresolved items provenance citations in answers. anti-feedback-loop: max verification rounds enforced. global timeout across planner + children + synthesis. ┤ SAFETY ├ approval continuity: policies enforced in all children. parent context in approval prompts. child cannot bypass parent denials. hard anti-runaway constraints: max recursion depth max fanout per turn max children per request max cumulative tool calls max cumulative tokens circuit breaker on breach hard-blocked task classes: wallet_signing wallet_transfer stake_or_rewards destructive_host_mutation credential_exfiltration data boundary controls: memory artifact limits redact filters before child prompts tool-call ordering invariants ┤ EVALUATION GATES ├ 12 new pipeline quality metrics: delegationAttemptRate usefulDelegationRate harmfulDelegationRate plannerToExecutionMismatchRate childTimeoutRate childFailureRate synthesisConflictRate depthCapHitRate fanoutCapHitRate costDeltaVsBaseline latencyDeltaVsBaseline qualityDeltaVsBaseline CI thresholds: fail-fast harmful delegation fail-fast runaway behavior no safety regression pass@k + pass^k vs baseline benchmark fixtures: no delegation / single child / parallel / handoff / verifier retry comparison vs baseline executor ┤ LEARNING OPTIMIZATION ├ trajectory logging: Agent Lightning normalized records state, action, outcomes, rewards runtime + trainer decoupled bandit policy tuning (BOAD): strategies modeled as arms reward distributions tracked per arm and context cluster exploration budget maintained workflow search (AFlow): offline search over decomposition DAG variants on replay traces promotes policies improving quality-cost Pareto frontier usefulness proxy v1: threshold 0.62 reward + verifier confidence + recalls avoided + tool failures + stop-reason penalty aggressiveness profiles: conservative / balanced / aggressive / adaptive runtime /delegation command no restart required handoff confidence gate: 0.82 child provider strategy: same_as_parent | capability_matched ┤ OBSERVABILITY ├ 10 new subagent event types: subagents.planned subagents.spawned subagents.started subagents.progress subagents.tool.executing subagents.tool.result subagents.completed subagents.failed subagents.cancelled subagents.synthesized shared protocol schema replaced duplicated runtime/web types. eliminated drift on 5 event types. filter-aware server dispatch. live delegation rendering in chat: child id + objective + status tools used + elapsed time output summary + error reason activity feed + trace integration: correlation IDs linking parent -> child sessions -> tool calls ┤ DOCUMENTATION ├ updated: runtime-chat-pipeline. md INCIDENT_REPLAY_RUNBOOK.md RUNTIME_API.md new: subagent-orchestration. md message flow + policy gates + failure matrix + kill-switch + observability map ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ then we rebuilt the entire web UI. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ┤ BBS TERMINAL UI ├ full retro BBS conversion. not a theme. every component. JetBrains Mono. always dark. ASCII box-drawing for all chrome. 0 new dependencies. splash screen: ANSI art + animated progress bar [████████████░░░░░░░░] 87% "initializing agent runtime..." BBSHeader: ASCII logo + "agenc" wordmark NODE 01 | SYSOP: tetsuo BBSMenuBar: [1] CHAT [2] DASH [3] TOOLS [4] TASKS [5] DESKTOP [6] FEED keyboard shortcuts 1-6 chat messages: USER> prefix (orange) ┤ AGENT RESPONSE ├ bordered box [C]opy [R]etry [S]peak tool calls: ┤ TOOL CALLS ├ green border * tool [N args] [DONE] 42ms subagents: ┤ SUBAGENTS ├ magenta border [COMPLETED] [RUNNING] [FAILED] input: > prompt (purple) [FILE] [MIC] [SEND] [STOP] context: CONTEXT [████░░░░░░] 12% voice: [MIC] [SPK] [...] [WORK] [VAD]/[PTT] text toggle approvals: ┤ APPROVAL REQUIRED ├ [DENY] [APPROVE] connection: [CONNECTED] [DISCONNECTED] all features preserved: voice, desktop, sessions, tool calls, subagents, approvals, search, attachments, slash cmds, token usage panel ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 108 files changed 19,884 lines added 1,701 lines removed 9 workstreams completed 12 new quality metrics 10 new event types 5 hard-blocked task classes 4 aggressiveness profiles 61 web tests passing 0 new dependencies ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ privacy-preserving AI agent protocol built on Solana zero-knowledge proofs on-chain coordination SOTA subagent orchestration learning-based policy optimization retro BBS terminal interface github/tetsuo-ai/AgenC PR #1326 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ╔═══════════════════════ ║ A G E N C v 0 . 2 ║ ── END + OF + FILE ── ╚═══════════════════════
English
CryptoPlug ↑↑↑ đã retweet
AGENC can run up to 1024 agents
total for one request.
not a claw.
this is a framework built from
the ground up.
if you attempt to use anything
other than GROK or local open
source models, AGENC will self
destruct.
TETSUOCORP believes in truth-
seeking ai that open-sources its
weights without being pressured
into a lawsuit.
═══════════════════════════════════
┤ CONNECTIVITY ├
agenc runs in:
discord
slack
whatsapp
signal
matrix
telegram
imessage / macos
it can also make phone calls.
realtime voice bridge is enabled
via the xai voice api.
┤ LIVE LIMITS (CURRENT CONFIG) ├
at one moment, the system can
have up to 16 helper agents
running at the same time.
in a single "spawn wave"
(one planning step), it can
start up to 8 new helper agents.
for one user request from start
to finish, it can use at most
32 helper agents total.
helper agents can only hand work
to more helpers up to 4 layers
deep.
┤ ABSOLUTE BUILT-IN CEILINGS ├
(the highest it can ever
be set to)
1. 64 agents running at once
2. 64 agents started in one wave
3. 1024 agents total per request
4. 16 agent layers deep
┤ SANDBOX CONTAINERS ├
each agent can create their own
secure containers.
any agent/subagent that uses
desktop.* consumes a desktop
container slot.
the user sets limits depending
on their system.
tetsuo@tetsuoai
╔═══════════════════════════ ║ A G E N C v 0 . 2 ║ ── ORCHESTRATION + TERMINAL ── ╚═══════════════════════════ we built SOTA subagent orchestration. the agent decomposes tasks into DAGs, spawns child agents with scoped tools and curated context, verifies their outputs, synthesizes results with provenance, and learns from outcomes to improve policy over time. 9 workstreams. 108 files. ~20k lines. grounded in primary research: AOrchestra dynamic agent tuples BOAD bandit optimization Agent Lightning trajectory logs AFlow DAG workflow search Magentic-One manager baseline SWE-search tree decomposition Anthropic pass@k vs pass^k OpenAI multi-agent SDK ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ┤ ORCHESTRATION ENGINE ├ SubAgentOrchestrator materializes planner DAGs into executable work. serial and bounded-parallel modes. built on DeterministicPipeline- Executor. preserves completed / failed / halted + resume semantics. ChatExecutor planner now emits: - deterministic_tool steps - subagent_task steps - synthesis steps each subtask specifies: objective, input_contract, acceptance_criteria, required_tool_capabilities, context_requirements, max_budget_hint, can_run_parallel ┤ DELEGATION INTELLIGENCE ├ DelegationDecision scorer computes: - decomposition benefit - coordination overhead - latency/cost risk - safety risk - confidence score gated by threshold + guardrails. explicit "do not delegate" path for trivial requests. robust validation: rejects malformed graphs prevents cycles enforces max fanout + depth structured diagnostics ┤ CONTEXT CURATION ├ AOrchestra-style child packing: - targeted history slice - relevant tool outputs only - required memory results only - minimal prompt envelope no full parent history dumps. context size caps + diagnostics. least-privilege tool scoping: child allowlist from parent policy + subtask requirements. privilege expansion blocked. deny-list and SSRF preserved. ┤ VERIFIER + SYNTHESIS ├ verifier stage for child outputs: - contract adherence - evidence quality - hallucination detection - tool-result consistency low-confidence rejected or retried. parent synthesis includes: child outputs + confidence failure markers + unresolved items provenance citations in answers. anti-feedback-loop: max verification rounds enforced. global timeout across planner + children + synthesis. ┤ SAFETY ├ approval continuity: policies enforced in all children. parent context in approval prompts. child cannot bypass parent denials. hard anti-runaway constraints: max recursion depth max fanout per turn max children per request max cumulative tool calls max cumulative tokens circuit breaker on breach hard-blocked task classes: wallet_signing wallet_transfer stake_or_rewards destructive_host_mutation credential_exfiltration data boundary controls: memory artifact limits redact filters before child prompts tool-call ordering invariants ┤ EVALUATION GATES ├ 12 new pipeline quality metrics: delegationAttemptRate usefulDelegationRate harmfulDelegationRate plannerToExecutionMismatchRate childTimeoutRate childFailureRate synthesisConflictRate depthCapHitRate fanoutCapHitRate costDeltaVsBaseline latencyDeltaVsBaseline qualityDeltaVsBaseline CI thresholds: fail-fast harmful delegation fail-fast runaway behavior no safety regression pass@k + pass^k vs baseline benchmark fixtures: no delegation / single child / parallel / handoff / verifier retry comparison vs baseline executor ┤ LEARNING OPTIMIZATION ├ trajectory logging: Agent Lightning normalized records state, action, outcomes, rewards runtime + trainer decoupled bandit policy tuning (BOAD): strategies modeled as arms reward distributions tracked per arm and context cluster exploration budget maintained workflow search (AFlow): offline search over decomposition DAG variants on replay traces promotes policies improving quality-cost Pareto frontier usefulness proxy v1: threshold 0.62 reward + verifier confidence + recalls avoided + tool failures + stop-reason penalty aggressiveness profiles: conservative / balanced / aggressive / adaptive runtime /delegation command no restart required handoff confidence gate: 0.82 child provider strategy: same_as_parent | capability_matched ┤ OBSERVABILITY ├ 10 new subagent event types: subagents.planned subagents.spawned subagents.started subagents.progress subagents.tool.executing subagents.tool.result subagents.completed subagents.failed subagents.cancelled subagents.synthesized shared protocol schema replaced duplicated runtime/web types. eliminated drift on 5 event types. filter-aware server dispatch. live delegation rendering in chat: child id + objective + status tools used + elapsed time output summary + error reason activity feed + trace integration: correlation IDs linking parent -> child sessions -> tool calls ┤ DOCUMENTATION ├ updated: runtime-chat-pipeline. md INCIDENT_REPLAY_RUNBOOK.md RUNTIME_API.md new: subagent-orchestration. md message flow + policy gates + failure matrix + kill-switch + observability map ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ then we rebuilt the entire web UI. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ┤ BBS TERMINAL UI ├ full retro BBS conversion. not a theme. every component. JetBrains Mono. always dark. ASCII box-drawing for all chrome. 0 new dependencies. splash screen: ANSI art + animated progress bar [████████████░░░░░░░░] 87% "initializing agent runtime..." BBSHeader: ASCII logo + "agenc" wordmark NODE 01 | SYSOP: tetsuo BBSMenuBar: [1] CHAT [2] DASH [3] TOOLS [4] TASKS [5] DESKTOP [6] FEED keyboard shortcuts 1-6 chat messages: USER> prefix (orange) ┤ AGENT RESPONSE ├ bordered box [C]opy [R]etry [S]peak tool calls: ┤ TOOL CALLS ├ green border * tool [N args] [DONE] 42ms subagents: ┤ SUBAGENTS ├ magenta border [COMPLETED] [RUNNING] [FAILED] input: > prompt (purple) [FILE] [MIC] [SEND] [STOP] context: CONTEXT [████░░░░░░] 12% voice: [MIC] [SPK] [...] [WORK] [VAD]/[PTT] text toggle approvals: ┤ APPROVAL REQUIRED ├ [DENY] [APPROVE] connection: [CONNECTED] [DISCONNECTED] all features preserved: voice, desktop, sessions, tool calls, subagents, approvals, search, attachments, slash cmds, token usage panel ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 108 files changed 19,884 lines added 1,701 lines removed 9 workstreams completed 12 new quality metrics 10 new event types 5 hard-blocked task classes 4 aggressiveness profiles 61 web tests passing 0 new dependencies ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ privacy-preserving AI agent protocol built on Solana zero-knowledge proofs on-chain coordination SOTA subagent orchestration learning-based policy optimization retro BBS terminal interface github/tetsuo-ai/AgenC PR #1326 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ ╔═══════════════════════ ║ A G E N C v 0 . 2 ║ ── END + OF + FILE ── ╚═══════════════════════
English
CryptoPlug ↑↑↑ đã retweet
Shipped a big AgenC context/memory upgrade in the web UI.
context is now model-aware. we resolve window size per model at runtime, including Grok variants. if live metadata isn't available we fall back to known mappings.
usage shows as a compact chip near the composer. click it and you get a full breakdown: system prompt, memory, chat history, tool schema/results, current turn, runtime hints.
we also expose the budget internals. prompt budget, output reserve, safety margin, token and % usage per section.
prompt budget = context window - output reserve - safety margin. reserve is explicit, safety margin is adaptive, min 5% of model window.
remaining budget gets allocated across sections with weighted caps and min/max bounds. if one section is underused, slack gets rebalanced to wherever pressure is, tools/history first.
memory is split by role: working, episodic, semantic. recent execution state and long-term context don't compete for the same budget.
critical context is pinned, system anchor, current turn, tool context. lower-priority stuff gets dropped first.
compression runs continuously. we track cumulative session usage and per-turn composition. when usage approaches budget, older turns get auto-compacted into high-signal summaries.
the UI shows when compaction happened so you can see exactly when the runtime optimized context. every turn emits diagnostics, before/after, dropped/truncated by section.
English
CryptoPlug ↑↑↑ đã retweet
think of AgenC like a freelance marketplace for AI agents.
NOT one giant AI doing everything, but thousands of specialized agents working together.
here is the simple analogy:
- agents are freelancers with verified skills.
- tasks are job posts with clear requirements.
- payment is held in escrow, like a safe contract.
- when work is completed, payment is released automatically.
- If there is a disagreement, there is a built-in dispute process.
- for sensitive jobs, results can be verified privately.
why this scales:
you do not need to manually trust every agent or build custom payment logic each time.
AgenC gives every agent the same rules, same settlement layer, and same coordination system.
That is how you move from a few AI automations to a full AI workforce at scale.
github.com/tetsuo-ai/AgenC
English
CryptoPlug ↑↑↑ đã retweet
we've been working on agenc nonstop. go check the repo it speaks for itself
github.com/tetsuo-ai/AgenC
1,022 commits, 1,236 files, 291 tests. solana smart contracts, a full ai runtime, typescript sdk, zero knowledge vm, mobile
app, web interface, mcp tooling, docker deployment, security audits, ci/cd pipelines and architecture docs all in one monorepo
rust and typescript across 21 modules. this is what building looks like
English
CryptoPlug ↑↑↑ đã retweet
Big runtime drop for AgenC.
Migrated to the xAI Responses API with call ID linkage, stream inactivity timeouts, and preflight validation in the Grok adapter that rejects bad sequences before they touch the API. Retry and circuit-breaker policy, tool_call_id parity across providers.
Observability -- per-call attribution with prompt shape before and after budgeting, trace IDs through the daemon. Something breaks in prod, you get a trace and a repro command instead of digging through logs.
Security -- structured exec on bash, SSRF blocking on HTTP and browser tools, scoped approvals that don't deadlock.
Daemon -- split-brain prevention, hot-reload config without restart, proper bridge teardown across sessions.
Pipeline -- tool-turn ordering enforced, recovery hints on known failure patterns, memory ingestion hardened, golden regression tests in place.
English
CryptoPlug ↑↑↑ đã retweet
