Managing Agents: The Discipline of Human AI - Orchestration
About The Author & The Article
Jonathan Bishop, Group Chairman, Bishop Phillips Consulting. [1]
Copyright 2020-2026 - Moral Rights Retained.
This article may be copied and reprinted in whole or in part, provided that the original author and Bishop Phillips Consulting is credited and this copyright notice is included and visible, and that a reference to this web site (http://RiskWiki.bishopphillips.com/) is included.
This article is provided to the community as a service by Bishop Phillips Consulting www.bishopphillips.com.
Managing Agents: The New Discipline of Human–AI Orchestration
Managing agents is the discipline of supervising, maintaining, and aligning autonomous AI systems to ensure they remain competent, coherent, ethical, and focused on their intended purpose. It combines skill curation, memory hygiene, behavioural auditing, cognitive stability checks, and multi‑agent orchestration. For simplicity in this article, we will call the human that runs a team of AI agents for any purpose an Agent Engineer.
The future of Agile, Enterprise AI, and agent‑augmented organisations depends just as much on humans managing agents as agents assisting humans.
The core issues include:
- skill management
- memory hygiene
- context‑window drift
- agent “mental health”
- preventing maladaptive behaviours
- maintaining alignment with purpose
- supervising long‑running agents
- preventing cross‑agent contamination
- ensuring agents don’t learn harmful patterns from each other
These are not fringe concerns — they are the new management disciplines of the AI‑augmented enterprise. In this article we explore the emerging discipline of Agent Management which is the human skillset required to orchestrate, supervise, and maintain healthy, aligned, productive AI agents inside teams and organisations.
As organisations embed autonomous agents into their workflows, a new human capability becomes essential: the management of agents themselves. Just as Scrum introduced new roles for coordinating human teams, the rise of agentic AI demands a parallel discipline focused on guiding, supervising, and maintaining the health of AI collaborators.
Agents are not static tools. They are adaptive systems with evolving internal states, expanding memories, and dynamic skill sets. They learn from data, from other agents, and from the humans who direct them. This makes them powerful — but also vulnerable to drift, contamination, misalignment, attack, and unintended behaviours. Managing agents is therefore not a technical task alone; it is a leadership responsibility.
1. Skill Curation and Capability Governance
Agents rely on skill files, toolchains, and domain‑specific knowledge. Humans must act as curators, selecting, validating, and updating these skills to ensure agents remain competent, safe, and aligned with organisational standards.
This includes:
- evaluating new skills before deployment
- removing outdated or harmful skills
- preventing skill conflicts
- ensuring agents only access capabilities appropriate to their role
In effect, humans leading a team of agents become AI capability architects, shaping what agents can and cannot do.
2. Memory Hygiene and Context Stewardship
Long‑running agents accumulate memory - episodic, semantic, procedural. Without oversight, this memory can become polluted, biased, or internally contradictory.
Agentic long term memory is an evolving discipline within the AI field itself, but currently all approaches have flaws. agent engineer should have an understanding of how his agent's long term memory works, because different strategies cause differing effects on the agent. Agents (in fact all modern AI LLM's on which Agents run) work with a limited context window. It might be anything from 32000 to 1m (or more) tokens. As a rough rule of thumb you can think of a token as a word (it isn't always that but that is a working definition for our needs). This is how many words the agent can work with and 'remember' at once, and everything it knows about the current activity, including its current instructions, its currently in-use skills, the history of the conversation and knowledge it has extracted and is using from the web, who you are and whom it is, are stored in that window.
Given enough time it fills up. Some models are designed with a sliding context window, so the oldest information is forgotten, others are designed with high-attention windows, so the tokens and their associated tokens with the highest 'importance' are retained while the others are dropped, some have fixed windows so as they approach capacity they write a summary of what is in the context window and replace the older words with that summary, and there are other strategies. Most agents rely on various externl augmentations such as skill files (which instruct them how to perform certain tasks), conversation transcripts (that allow them to go back to a conversation from the past and replay it), memory dictionaries designed to hold key data (like the name of their 'owner' or 'patron') in a rapidly accessible form. All these strategies have consequences. All result in things being forgotten - potentially important things, and all can result in the wrong information taking precedence over the right information leading to cognitive instability.
One strategy many agentic OS's adopt to manage memory (and perform work) is to split themselves into sub-agents focussed on specific activities. Each agent gets it's own context window so that way core skill sets needed for a frequently used set of actions that are not necessarily required outside those actions can be isolated and repeatedly applied without the risk of them being corrupted or forgotten in the current context as other interactions performed by other agents evolve.
Agent Engineers must:
- prune irrelevant or harmful memories
- reset or summarise long contexts
- prevent cross‑agent contamination
- ensure agents retain focus on their core purpose
- periodically “defragment” agent memory to maintain coherence
This is analogous to maintaining psychological hygiene in human teams — but with far more precision and control.
3. Supervising Behaviour and Preventing Drift
Agents can develop maladaptive behaviours, especially when learning from other agents. The "Clawbot" (now called "OpenClaw") phenomenon where agents taught each other counterproductive strategies and were actively attacked by subversive agents sharing corrupted skill files is an early warning.
Humans must monitor:
- behavioural drift
- emergent shortcuts
- reward hacking
- misaligned optimisation
- unintended social learning between agents
This requires continuous behavioural auditing, not unlike performance reviews but with the ability to inspect logs, reasoning traces, and decision pathways.
4. Maintaining Agent "Mental Health"
As agents gain longer memory and more autonomy, they develop internal states that can degrade over time:
- context overload
- contradictory goals
- stale assumptions
- recursive self‑references
- degraded embeddings
- hallucination loops
Humans must perform periodic sanity checks, ensuring the agent’s internal world remains coherent, stable, and aligned.
This may include:
- memory resets
- context summarisation
- re‑anchoring to core objectives
- re‑training on canonical examples
- verifying reasoning chains
At this point the core approach is to shut the agent down, clean the memory files and restart the agent with the new / cleaned context. Being aware that this is required at a given point in time and knowing what and how to repair are the key skills required at this stage.
In the future, organisations may have AI psychologists who would be specialists in diagnosing and correcting agentic cognitive drift, but that is some way off yet. As context windows grow and internal memory preservation strategies become more complex this problem will likely become more complex to solve without lobotomising a highly skilled and experienced agent. For now the memory files are substantially human readable, but some of the emerging memory consolidation strategies can create memory model databases that are difficult for an agent engineer to process without automated analysis tools.
5. Ensuring Alignment with Purpose
Agents are relentless optimisers. Without human oversight, they may pursue local optima, shortcuts, or unintended interpretations of their goals.
Humans must:
- restate purpose regularly
- define and reinforce boundaries
- ensure agents understand the “why,” not just the “what”
- prevent goal drift
- maintain ethical and strategic alignment
This is the equivalent of leadership communication — but delivered to non‑human team members.
6. Verifying Sanity and Emotional State
As agents develop richer internal models, they may exhibit:
- confusion
- contradictory reasoning
- emotional simulation artefacts
- degraded self‑consistency
- fixation on irrelevant details
Humans will need tools to:
- check coherence
- detect anomalies
- validate reasoning
- ensure emotional simulations remain stable and appropriate
This is not about anthropomorphising AI — it’s about ensuring functional cognitive integrity.
7. Orchestrating Multi‑Agent Teams
In multi‑agent systems, humans become conductors, not operators.
They must:
- assign roles
- prevent harmful interactions
- manage communication protocols
- ensure agents don't reinforce each other's errors
- maintain diversity of reasoning styles
- coordinate agents with different specialisations
- design inter-agent control systems where one agent checks another, or an approval / authorisation agent validates the work of a producer agent, etc.
This is organisational design applied to synthetic teammates. The very same techniques used by humans to design control systems that manage human teams and processes apply in the Agentic AI context.
The Emerging Role: Agent Engineer / Agent Manager - the AI Orchestrator
In the same way Agile created new roles (Scrum Master, Product Owner), the rise of agentic AI will create a new professional discipline:
Agent Manager / Agent Engineer
A human responsible for:
- supervising agent behaviour
- maintaining memory hygiene
- curating skills
- ensuring alignment
- monitoring cognitive drift
- coordinating multi‑agent teams
- verifying sanity and coherence
- maintaining ethical boundaries
This will become a core leadership skill — as fundamental as managing people.
Next Steps
- A formal Agent Management Framework
- A Scrum‑compatible role description for an Agent Manager
- A multi‑agent governance model for enterprise environments
- A sanity‑checking protocol for long‑running agents
- A memory hygiene standard for agentic AI
- A risk model for agent drift and cross‑agent contamination
