Perplexity Ships "Teammates" — Agents That Don't Clock Out

Perplexity Ships "Teammates" — Agents That Don't Clock Out

Perplexity launched Teammates on November 5, persistent agents for Max subscribers ($200/mo) that remember context, hold projects, and work agent-to-agent without a human re-prompting them every five minutes [1]. CEO Aravind Srinivas didn't bury the lede: Perplexity has been running roughly 10,000 internal "teammates" for 30 days — about 20x their actual headcount — with median employees now logging 7x the agent-hours they were in September, and one in five employees pushing their agents over 40 hours a week [2][3].

What are these agents actually doing? Not toy tasks — bug triage, QA, security monitoring, A/B testing, growth marketing. This is a company running its own internal operations as a testbed for AI-native workflows, and the numbers suggest it's sticking.

Srinivas has been vocal on X about agents prompting other agents and using git as the canonical source of truth rather than a human inbox. That's a meaningful tell: when the "manager" in a workflow is increasingly another agent, your org chart starts looking less like reporting lines and more like orchestration graphs.

Exo: Agents That Get Cheaper and Better While They Work

A new arXiv paper out of UC Berkeley (submitted October 30) introduces Exo, a runtime architecture that lets agents safely edit their own code and policies mid-task — in-harness learning, not retraining on some external dataset after the fact [1][2]. The results are the kind of numbers that get CFOs' attention: 35-61% cost reductions on AppWorld and SREGym benchmarks at comparable correctness, with SREGym diagnosis time cut 35% and AppWorld task time down 13% [1].

This is a different flavor of progress than "bigger model, better benchmark." Exo is about agents that improve their own operating procedure as they run, inside guardrails designed to keep that self-editing safe. On the Continual Learning Bench, it shaved 14% off costs even after accounting for the overhead of learning itself.

The GitHub repo is already drawing attention from builders who want self-improving harnesses without the usual retraining pipeline tax [3]. If this generalizes beyond benchmarks, it's a preview of agents that get measurably cheaper to run the longer you use them — the opposite of how software usually ages.

Perplexity's "Ultra" Effort Mode Formalizes the Orchestration Layer

Perplexity quietly updated its Agent API documentation this month to add an "ultra" effort level — built for long-running, high-coverage research and data tasks on Claude Opus 5.5 — alongside existing fast/low/medium/high/xhigh tiers [1][2]. Each level bundles a specific combination of models, tools, and prompts, letting developers trade latency for depth and tool use explicitly rather than guessing at prompt engineering tricks.

This is a small change with a big implication: effort level is becoming a first-class configuration parameter, the same way you'd set memory or timeout limits on a cloud function. Combined with Teammates, it gives enterprises a dial for how much autonomy and compute to hand an agent for a given job.

Srinivas's X commentary on agent-to-agent prompting and git-as-truth-source lines up directly with this release — the API is clearly being built to support fleets of agents coordinating each other, not single-shot queries [3].