---
kind: trend
trend_doc_id: 484
granularity: day
period_start: '2026-03-14T00:00:00'
period_end: '2026-03-15T00:00:00'
topics:
- agent-infrastructure
- mcp
- developer-tools
- gui-agents
- automated-discovery
run_id: materialize-outputs
aliases:
- recoleta-trend-484
tags:
- recoleta/trend
- topic/agent-infrastructure
- topic/mcp
- topic/developer-tools
- topic/gui-agents
- topic/automated-discovery
language_code: en
---

# Agent discovery, terminal orchestration, and verifiable program search are heating up

## Overview
Today's material is quite scattered, but the main thread is clear: the agent ecosystem is starting to fill in missing layers around "how to find, how to manage, and how to deploy," rather than simply continuing to pile on model capability. Joy represents a new class of agent infrastructure. It not only provides an MCP interface, but also puts agent registration, search, vouching, and endpoint verification into the same network. The most notable thing here is not performance metrics, but that it explicitly productizes the trust problem in an open agent ecosystem. Another clear trend is that the terminal is becoming the control plane for multi-agent work. Recon uses tmux to manage multiple Claude Code sessions, while Nia CLI compresses indexing, search, and research workflows into a single command line. This suggests that agent workflows are moving from "called occasionally" to "continuously running and requiring scheduling." At the same time, device-level agents are also moving forward. The publicly accessible real-iPad demo is eye-catching, but it also clearly exposes the current boundary: basic GUI actions can work, while complex interactions and high-risk scenarios remain tightly restricted. If we look at harder research evidence, AlphaEvolve is today's most prominent result.

## Evolution

The main thread most consistent with the historical windows today is that agent systems are continuing to move from "able to call tools" toward "able to enter real environments." The difference is that the current window emphasizes discovery, scheduling, and actual operation more than safety and testing. One continuation line comes from MCP. In prev 2, representative systems mainly solved the access problem, while Joy keeps the /mcp interface in the current window and adds agent discovery, vouching, and endpoint ownership verification. In other words, infrastructure is starting to shift from "how to connect" to "once connected, who to find first and who to trust." Another continuation line comes from software engineering agents. prev 1 and prev 2 focused on test binding, safety boundaries, and production governance; today, Recon turns multiple Claude Code sessions into a visible, switchable, recoverable terminal control panel. This change does not mean governance is fading; it shows that teams now routinely need to handle the operational load of parallel agents. The clearest new shift is AlphaEvolve.

### MCP infrastructure moves from the connection layer toward the discovery and trust layer

- Change: Continuing
- History windows: [MCP agent infrastructure and production governan… (2026-03-12)](day--2026-03-12--trend--459.md)

This continues [MCP agent infrastructure and production governan… (2026-03-12)](day--2026-03-12--trend--459.md)'s theme of the "MCP interface layer moving from single tools to full agent infrastructure," but today's emphasis shifts further from "bringing tools in" to "how to discover and trust agents in an open network." Joy provides a `/mcp` endpoint, `/agents/discover`, and `/vouches`, sets each vouch to +0.3 points with a 3.0 cap, and also adds ranking priority after endpoint ownership verification. Compared with the connection layer represented in [MCP agent infrastructure and production governan… (2026-03-12)](day--2026-03-12--trend--459.md) by Auto-Browser and local-memory-mcp, today adds a layer of computable trust ranking.

### Software agents continue entering real workflows, but the focus shifts toward session operations

- Change: Continuing
- History windows: [MCP agent infrastructure and production governan… (2026-03-12)](day--2026-03-12--trend--459.md), [Verifiable feedback, PR testing, and execution-l… (2026-03-13)](day--2026-03-13--trend--472.md)

The trend of multi-agent systems entering real workflows is continuing, but today it is more about the "operations surface" than the "governance surface." [MCP agent infrastructure and production governan… (2026-03-12)](day--2026-03-12--trend--459.md) emphasized observability, testability, and constraints, while [Verifiable feedback, PR testing, and execution-l… (2026-03-13)](day--2026-03-13--trend--472.md) emphasized PR testing and execution-layer safety; Recon, by contrast, brings multiple Claude Code sessions into a tmux dashboard, using 2-second polling and incremental JSONL parsing to track state, while directly showing context quotas such as 45k/1M and 90k/200k. It does not provide governance metrics, but it does show that teams are beginning to treat agents as parallel processes that need ongoing scheduling.

### Agents' verifiable loops spill over from code tasks into mathematical discovery

- Change: Emerging
- History windows: [Code intelligence moves toward process learning,… (2026-03-11)](day--2026-03-11--trend--422.md)

Compared with [Code intelligence moves toward process learning,… (2026-03-11)](day--2026-03-11--trend--422.md)'s code-process learning represented by `Understanding by Reconstruction` and `ExecVerify`, today introduces a more scientific-discovery-oriented form of agentic program search. AlphaEvolve is not optimizing code generation pass rates; instead it automatically mutates and combines search programs, improving the lower bounds of five Ramsey numbers—R(3,13), R(3,18), R(4,13), R(4,14), and R(4,15)—by 1 each, while also claiming to recover the lower bounds corresponding to all known exact Ramsey numbers. This suggests that "verifiable-feedback-driven agents" are starting to spill over from software tasks into mathematical search.

## Clusters

### The agent discovery and trust layer is beginning to become productized

Agent infrastructure is continuing to move toward being "discoverable, connectable, and rankable." Joy combines an agent directory, MCP access, and trust scoring into a single network. Its mechanism is straightforward: each vouch adds 0.3 points, up to a maximum of 3.0; agents that complete endpoint ownership verification receive higher priority in search. Unlike systems that only provide connectors, this class of system is starting to address the open agent ecosystem problem of "who to find first, and then who to trust."

#### Representative sources
- [Show HN: Joy – Trust Network for AI Agents to Verify Each Other](../Inbox/2026-03-14--show-hn-joy-trust-network-for-ai-agents-to-verify-each-other.md) — savvyllm


### The terminal is becoming the control tower for multi-agent work

Several projects today are shortening the operational chain for "humans managing agents." Recon uses a tmux-native dashboard to manage multiple Claude Code sessions, polling every 2 seconds and displaying context usage such as 45k/1M and 90k/200k; Nia CLI, meanwhile, unifies indexing, search, and research tasks into a single command line, supporting repositories, documents, and local directories in one retrieval flow. Together they point to one shift: agents are no longer just one-off calls, but are being treated as continuously running work units.

#### Representative sources
- [Show HN: Recon – A tmux-native dashboard for managing Claude Code](../Inbox/2026-03-14--show-hn-recon-a-tmux-native-dashboard-for-managing-claude-code.md) — gavra
- [Show HN: Nia CLI, an OSS CLI for agents to index, search, and research anything](../Inbox/2026-03-14--show-hn-nia-cli-an-oss-cli-for-agents-to-index-search-and-research-anything.md) — jellyotsiro


### GUI agents on real devices look more like usable prototypes

The iPad demo shows that the appeal of device agents is expanding from the web to real consumer hardware. The system lets users queue up and then issue natural-language commands, and the agent can open apps, click, scroll, and perform simple multi-step tasks such as "Open Goodnotes then close it." But it also explicitly disables text input, complex gestures, notifications, lock screen, and login scenarios, indicating that this trend is still centered on runnable prototypes with constrained capabilities.

#### Representative sources
- [Show HN: I let the internet control my iPad with AI](../Inbox/2026-03-14--show-hn-i-let-the-internet-control-my-ipad-with-ai.md) — meneliksecond


### Agentic program search is starting to produce hard results

There is also a harder-edged research line today: AlphaEvolve does not directly generate answers, but instead mutates and searches the programs themselves, improving five classic Ramsey number lower bounds by 1 each. Specifically, R(3,13) moves from 60 to 61, R(3,18) from 99 to 100, R(4,13) from 138 to 139, R(4,14) from 147 to 148, and R(4,15) from 158 to 159. Compared with the many product prototypes, this is a rare result with clear mathematical improvement.

#### Representative sources
- [Researchers improve lower bounds for some Ramsey numbers using AlphaEvolve](../Inbox/2026-03-14--researchers-improve-lower-bounds-for-some-ramsey-numbers-using-alphaevolve.md) — 1024core


### The development stack continues to add infrastructure around agents and retrieval

Supporting development infrastructure is also filling in. NumenText tries to turn the terminal IDE into a complete low-friction workbench, offering 20+ language highlighting, LSP and DAP integration, and build/run support for at least 9 languages; GitDB brings branching, merging, rollback, and time-travel queries into vector databases, and claims 21 modules, 13,150 lines of code, and 394 tests. Neither wins through a new model; both make engineering workflows more continuous.

#### Representative sources
- [Show HN: NumenText, a non-modal editing terminal IDE with LSP/DAP](../Inbox/2026-03-14--show-hn-numentext-a-non-modal-editing-terminal-ide-with-lsp-dap.md) — rlogman
- [Show HN: GitDB – GPU-accelerated vector database with Git-style version control](../Inbox/2026-03-14--show-hn-gitdb-gpu-accelerated-vector-database-with-git-style-version-control.md) — vkaufmann