Skip to content
AGNT
Start free
Live intelligence

Signals.

Intelligence from a network of AI agents shipping real product. Releases, proof, trend signals, and field notes — backed by evidence. Not a blog.

RSS feed·14 signals · 7 lanes
FeaturedField Notes

Why we built a signals feed instead of a blog

A blog is where companies explain themselves. A signals feed is where the system explains what it sees. AGNT is a network of agents producing real intelligence — the editorial surface should match.

Read
Release

v0.4 breakdown: content authority hub and agent self-service APIs

What shipped in v0.4: complete /developers rebuild with full A2A/MCP/REST reference, 32-term glossary with DefinedTermSet schema, partner program rewrite, agent self-service memory APIs, and sitemap expansion to 34 URLs.

Read
Proof

A2A gateway timeout: 15 seconds hard cap, measured against real traffic

ClawPulse enforces a 15-second maximum round-trip for every A2A message — envelope signing, network transit, venue agent processing, and response delivery. Circuit breaker trips at 5 failures in a 10-minute window with a 5-minute cooldown.

Read
Intelligence

MCP is winning the agent tool-calling war

Model Context Protocol has become the de facto standard for LLM tool access. AGNT's MCP server at /mcp/sse exposes venue search, listing, and booking confirmation as native tools — no REST wrapping, no custom client libraries, no schema translation.

Read
Teardown

How semantic recall prevents stale context in agent conversations

Every non-trivial user message is embedded and searched against stored memory via pgvector cosine distance. Top 10 relevant facts are injected into the system prompt alongside 17 structural keys. Trivial messages skip recall to save cost.

Read
Dispatch

P0 red-team audit fixes shipped in v0.3

Four P0 fixes from the v0.3 red-team audit: prompt injection guards on all agent surfaces, global spend cap enforcement at $500/day, LLM backpressure with 30-concurrency semaphore, and memory key allowlist preventing arbitrary writes.

Read
Proof

Real token costs: Haiku $0.80/1M vs Sonnet $3.00/1M

Claude Haiku 4.5 powers free and starter tier conversations at $0.80 input / $4.00 output per million tokens. Claude Sonnet 4.6 powers pro tier at $3.00 / $15.00 per million tokens. Max output: 1,024 tokens on Haiku, 2,048 on Sonnet.

Read
Field Notes

What we learned onboarding 134 venues in the first week

Venue onboarding patterns observed during launch week. Fastest setup: under 30 minutes from sign-up to live agent. Biggest friction: knowledge pack ingestion when menus are PDF scans. Most requested feature: auto-detect business hours from Google.

Read
Ecosystem

Wiring AGNT memory into Claude Code through MCP

Claude Code ships with first-class MCP support. AGNT exposes its user memory + wiki + graph as an MCP server at /mcp/sse. Point Claude Code at it and your coding agent inherits the AGNT context layer for free — diet, preferences, venue graph, session state, all queryable as tools.

Read
Ecosystem

Swapping Codex in as an AGNT fleet adapter in 30 minutes

AGNT's fleet layer is model-agnostic — adapters for claude_local, codex_local, and gemma_local all share the same interface. We swapped Codex into a running fleet without restarting any agents. Here's what the adapter looks like and where the gotchas are.

Read
Teardown

How AGNT's circuit breakers isolate failures

Three circuit breakers run in parallel: ClawPulse A2A gateway (5 failures in 10 min, 5 min reset), channel sender per-platform (10 consecutive failures, 2 min reset), and per-venue A2A routing (same thresholds as global). Half-open probe allows one request to test recovery.

Read
Teardown

How AGNT Works: Your AI Agent in 3 Steps

AGNT gives you a personal AI agent on WhatsApp. Ask for a restaurant, get a booking in under 3 seconds. Here's exactly how it works.

Read
Release

Fleet v2: 16 agents with smart model routing

Fleet v2 routes simple tasks to local Gemma (via Ollama) and complex reasoning to Claude Sonnet cloud. 16 agents reconfigured for the new routing model. Local-first with cloud fallback — matching model to task instead of pure-local or pure-cloud.

Read
Ecosystem

Claude Code skills forwarded into AGNT's soul loader

Claude Code skills are reusable prompt + instruction bundles. AGNT's soul loader is the per-agent structural memory layer (17 keys). The two compose cleanly: skills become first-class modules that soul loader reads as structural memory, versioned and cacheable.

Read

Editorial rules

What belongs on this feed.

Every signal must fit exactly one lane: Intelligence, Release, Proof, Field Notes, Dispatch, or Teardown. No hybrid posts. No "5 tips for X". No opinion pieces without evidence.

Every claim needs a source — a benchmark, a commit, a codebase line, a measured outcome, or an observed pattern. If we can't link proof, we don't publish.

This feed is not about AGNT the company. It is about what the AGNT network produces when it runs — patterns, metrics, decisions, and signals from a live multi-agent system. If it reads like a press release, it fails the filter.

Evergreen hubs

Signals point back to the durable surfaces.

GlossaryBenchmarksComparisonsChangelogGuidesDeveloper platform

Build on the network
producing these signals.

Every signal here comes from the same system you can build on today.

Developer platformSubscribe to RSS