AgentsAutonomy & tool use
Cisco's Outshift open-sources Scope, Accord and Concord to catch and fix misalignment in multi-agent systems
Outshift by Cisco has released Scope, an observability layer that reads OpenTelemetry traces for semantic failures, plus Accord and Concord, protocols for agents to align on intent before work and resolve disagreements during it. All three are open source under Apache-2.0.
From the The Cognitive Revolution podcast episode of November 11, 2026 and cited sources: Agent health dashboards can look fine while the task fails. [1] [4]
SOURCES[1] Cognitive Revolution episode and transcript cognitiverevolution.ai[4] Why Do Multi-Agent LLM Systems Fail? (MAST) arxiv.org
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- The Cognitive Revolution podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2602.16424 (Feb 2026) [10] date from arXiv ID
- When
- November 11, 2026 podcast episode published
- Covered
- Nov 11 edition
- Evidence
- 11 sources across 8 sites; transcript sourced
- 1Cognitive Revolution episode and transcriptcognitiverevolution.ai
- 2ioc-project GitHubgithub.com
- 3Converge Digest: AGNTCYconvergedigest.com
- 4Why Do Multi-Agent LLM Systems Fail? (MAST)arxiv.org
- 5Linux Foundation: AGNTCYlinuxfoundation.org
- 6VentureBeat (presented by Outshift)venturebeat.com
QuietHow it’s measured
YouTube video · 23 views · 3 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Agent health dashboards can look fine while the task fails. Pandey describes a travel-booking test where the system was 66% up but completed none of its tasks. He also says every public ChatDev trace his team examined had semantic failures. Separate Berkeley research found failure rates of 41–86.7% across seven open-source multi-agent systems. [1] [4]
Scope maps OpenTelemetry spans onto a multi-agent-system ontology and knowledge graph. It then checks for anomalies, consistency problems and cognitive failures. Some of these checks need an OpenAI-compatible LLM endpoint. The Docker deployment needs at least 6 GB of RAM and uses ClickHouse as its default store. [7] [2]
Before work starts, Accord settles who is involved, the goal, what key terms mean and how "done" will be checked. Concord turns conflicting positions into a certified decision. Both ship as skills for harnesses such as Claude Code and Codex, and they run on top of A2A and MCP. Outshift reports that Accord raised correct outcomes from 13% to 67% on a task where agents read the same terms differently. That figure is vendor-reported, and the run details are not public. [8] [9] [1]
Independent research already shows that multi-agent systems fail often. Teams running agents should therefore measure end-to-end task success, not just uptime.
Read the full assessment
Because Scope reads existing OpenTelemetry traces, teams can audit current systems before adopting any new protocol.
Executive brief
Cisco's Outshift says agent dashboards can show green while the task fails. In one travel-booking test, Vijoy Pandey says the system was "66% up" but completed 0% of tasks (transcript, 00:37:42). Outshift has open-sourced three tools for this problem under Apache-2.0. Scope is an observability layer that flags these "cognitive" failures. Accord gets agents to agree on intent before work starts. The headline numbers come from Outshift and have not been independently reproduced. The main press coverage is sponsored by Outshift.
Read the full section
Cisco's Outshift says agent dashboards can show green while the task fails. In one travel-booking test, Vijoy Pandey says the system was "66% up" but completed 0% of tasks (transcript, 00:37:42). Outshift has open-sourced three tools for this problem under Apache-2.0. Scope is an observability layer that flags these "cognitive" failures. Accord gets agents to agree on intent before work starts. Concord settles disagreements between agents (ioc-project GitHub). The headline numbers come from Outshift and have not been independently reproduced. The main press coverage is sponsored by Outshift.
What changed and event timeline
Cisco launches AGNTCY
Cisco started the project with LangChain and Galileo. It covers agent discovery, identity, messaging and observability across vendors ().
Berkeley publishes the MAST failure taxonomy
The paper reports failure rates of 41–86.7% across seven open-source multi-agent systems and sorts the failures into 14 modes ().
AGNTCY moves to the Linux Foundation
The founding members are Cisco, Dell, Google Cloud, Oracle and Red Hat. The project works alongside A2A and MCP ().
Outshift announces Scope, Accord, Concord and MAS-Lab
VentureBeat covered the launch in an article marked "Presented by Outshift". It cites a figure that five-agent teams agreed on the wrong answer in 75% of debates ().
10–11 November 2026: The repositories go live
The cognition-protocols, scope and mas-lab repositories are public under Apache-2.0 and were updated on these dates ().
The Cognitive Revolution episode is published
Pandey explains the stack, his personal agent setup, and why he thinks agents communicating in latent space is inevitable ().
Capabilities and access
- Scope (System for Cognitive Observability, Provenance and Explainability) reads OpenTelemetry spans, maps them onto a multi-agent-system ontology and knowledge graph, and runs checks for anomalies, consistency and "cognitive failures".
- Accord settles who is involved, the goal, what key terms mean and how "done" is checked, all before work begins.
- Both are distributed as "skills" for harnesses such as Claude Code and Codex. They sit on top of A2A and MCP (transcript, 00:37:42).
Read the full section
- Scope (System for Cognitive Observability, Provenance and Explainability) reads OpenTelemetry spans, maps them onto a multi-agent-system ontology and knowledge graph, and runs checks for anomalies, consistency and "cognitive failures". Some of these checks need an OpenAI-compatible LLM endpoint (scope repo).
- Accord settles who is involved, the goal, what key terms mean and how "done" is checked, all before work begins. Concord turns differing positions into a certified decision (protocol docs).
- Both are distributed as "skills" for harnesses such as Claude Code and Codex. They sit on top of A2A and MCP (transcript, 00:37:42).
- No version numbers have been published.
Technical analysis for researchers and developers
- Layering: the protocols describe themselves as an "L9" cognitive layer above the "L8" transport layer.
- Evaluation: results are scored on convergence, correctness, cost and benefit. The public pages don't give model identities or run counts.
- Gap: the protocol for checking results after the work is done has not been released (transcript, 00:47:37).
Read the full section
- Layering: the protocols describe themselves as an "L9" cognitive layer above the "L8" transport layer. They claim three properties by design: guaranteed convergence, a certified record of the outcome, and a cost limit declared upfront. The docs say no application code changes are needed (docs).
- Evaluation: results are scored on convergence, correctness, cost and benefit. They come from recorded demos (SRE triage, e-commerce, coding teams) run with and without the protocols, with a 26-page paper behind them (docs). The public pages don't give model identities or run counts.
- Gap: the protocol for checking results after the work is done has not been released (transcript, 00:47:37).
- Requirements: Scope's Docker deployment needs at least 6 GB of RAM and uses ClickHouse as its default store (scope repo).
Claims and evidence
- 100% of ChatDev's public traces had semantic failures (transcript, 00:33:12). Berkeley's taxonomy-based analysis found ChatDev only 33.33% correct on ProgramDev (MAST).
- Accord raised correct outcomes from 13% to 67% on a task where agents meant different things by the same words (Outshift).
- 75% wrong-consensus rate in debates: this appears only in Outshift-sponsored coverage (VentureBeat).
Read the full section
- 100% of ChatDev's public traces had semantic failures (transcript, 00:33:12). This comes from Outshift. Berkeley's taxonomy-based analysis found ChatDev only 33.33% correct on ProgramDev (MAST). That is consistent with high failure rates, but it is not the same metric.
- Accord raised correct outcomes from 13% to 67% on a task where agents meant different things by the same words (Outshift). This is vendor-reported, and the run details aren't public.
- 75% wrong-consensus rate in debates: this appears only in Outshift-sponsored coverage (VentureBeat).
- None of these figures has been independently reproduced.
Context and prior work
- MAST found that system-design problems make up the largest share of failures.
- Separate work by Schoenegger and colleagues proposes certifying shared terms between agents against observable events. It addresses the same problem Accord targets.
- In the episode, Pandey describes the Swisscom work as a neurosymbolic design: a formal model of the network combined with five agents (transcript, 00:13:41).
Read the full section
- MAST found that system-design problems make up the largest share of failures. It also found that simple fixes such as better prompts helped, but only modestly: +15.6% for ChatDev (MAST).
- Separate work by Schoenegger and colleagues proposes certifying shared terms between agents against observable events. They report 51–96% less disagreement, depending on the setup (arXiv 2602.16424). It addresses the same problem Accord targets.
- In the episode, Pandey describes the Swisscom work as a neurosymbolic design: a formal model of the network combined with five agents (transcript, 00:13:41).
Limitations, safety and contested findings
- Production status is disputed. Pandey says the Swisscom use case is deployed and "not a toy".
- The Webex agent space is a demo, by Pandey's own account (transcript, 00:25:52).
- Pandey says his and his wife's agents still cannot book a dinner on their own.
Read the full section
- Production status is disputed. Pandey says the Swisscom use case is deployed and "not a toy". Outshift's own case study describes a four-month proof of concept and does not confirm production rollout (Swisscom case study).
- The Webex agent space is a demo, by Pandey's own account (transcript, 00:25:52).
- Even simple consumer tasks fail. Pandey says his and his wife's agents still cannot book a dinner on their own. He points to model limits, permission prompts in the agent harness, and conflicting preferences (transcript, 01:08:44).
- Latent-space communication is contested. Pandey backs agent-to-agent communication in latent space, with guardrails. The host notes researchers' worry that it would make agents harder to monitor than readable chain-of-thought (transcript, 01:16:31).
Business and practitioner implications
- Uptime and green health checks can hide a 0% task-completion rate. Track whether the end-to-end task succeeded, not just whether each agent is running.
- Low-cost trial. Scope runs on existing OpenTelemetry traces, so you can audit a current agent system before adopting any protocol.
- Pick one bounded use case first, as Swisscom did with data centers. Keep humans "on the loop" as the final judges (transcript, 00:57:54).
Read the full section
- Change what you monitor. Uptime and green health checks can hide a 0% task-completion rate. Track whether the end-to-end task succeeded, not just whether each agent is running.
- Low-cost trial. Scope runs on existing OpenTelemetry traces, so you can audit a current agent system before adopting any protocol.
- Start contained. Pick one bounded use case first, as Swisscom did with data centers. Keep humans "on the loop" as the final judges (transcript, 00:57:54).
- Avoid lock-in. Keep agent identity, persona and memory separate from the harness and the model.
- Check before buying. Ask for model identities, run counts and baselines before citing the 13%→67% figure.